[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/mlp/ - Pony


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: CHAG.png (890 KB, 1280x1280)
890 KB PNG
/CHAG/ #180 — Alicorn Intelligence Edition

Welcome to Chatbot AI General #180, the thread for discussing and improving AI pony chatbots.

▶ MLP Bots
https://mlpchag.neocities.org
Spreadsheet (CAI bots + Old repository):
https://docs.google.com/spreadsheets/d/1J7BeqJVDS51cXF8Pgm2YZaFq-Z6ykSJT
SillyTavern website extension: https://github.com/MLPChag/SillyTavern-Chag-Search
Bunch of Character Cards (Descriptions/No Greetings):
https://docs.google.com/spreadsheets/d/1Y6LNOCqAZAWIX-OBEv55HjlzcEdpUh-XbKCpbpL6v5k
CAI bots converted to Tavern: https://files.catbox.moe/ckurq1.zip
bonziPONY: https://github.com/maresmaresmares/bonziPONY
!!!GALLERY!!!: https://drive.google.com/drive/u/2/folders/1Ao-h5HFGMPllSrzSBKM_BvGSiU9f0c2U

▶ How do I start?
1) Select a Frontend
2) Select an AI model
3) Select Jailbreak
4) Select bots
5) Lovemaking with AI mares!

Starting in this hobby can be confusing and difficult. If it’s your first time and you’re lost,
▶ ASK THE THREAD! ◀

Novice-to-advanced guide: https://rentry.org/onrms
https://rentry.org/sukino-findings

▶ SillyTavern (preferred frontend)
https://github.com/SillyTavern/SillyTavern
On Android: https://rentry.org/TermuxST
Pony TTS app supporting SillyTavern: https://drive.google.com/drive/folders/16Ss26VBmgzcSuTGzhaHqRuyVRceTf-YB

▶ More frontends:
Risu: https://risuai.xyz
Agnai: https://agnai.chat

▶ Locals
https://rentry.org/lunarmodelexperiments
>>>/g/lmg/
Mistral Nemo base model fine-tuned on fimfics: https://huggingface.co/Ada321/Nemo_Pony_2/tree/main

▶ Jailbreaks
MLP JB: https://rentry.org/znon7vxe
More JB and guides: https://rentry.org/jb-listing
Hypebots for Tavern: https://rentry.org/pn3hb

▶ Botmaking
Editors: https://agnai.chat/editor
Guides: https://rentry.org/meta_botmaking_list
Advanced: https://rentry.org/AdvancedCardWritingTricks

▶ Miscellaneous
Scrapped/Collected rentries: rentry.org/mrhd
Chag Arts: https://drive.google.com/drive/u/2/folders/1yfajin_hB5rW_jy6WyFjRaRfDA3EUGiJ
https://rentry.org/ChagMiscellaneous
https://rentry.org/ChagArchive

Previous thread: >>43209690

**News:**

- Deepseek V4 released https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf
- GPT-5.5 released https://openai.com/index/introducing-gpt-5-5
- Kimi 2.6 released https://www.kimi.com/blog/kimi-k2-6
- Claude Opus 4.7 released https://www.anthropic.com/news/claude-opus-4-7
>>
File: Anchor.png (552 KB, 1201x1200)
552 KB PNG
Anchor for bots, lorebooks, scenarios.
Previous >>43209692
>>
File: TAnchor.png (40 KB, 512x768)
40 KB PNG
Anchor for technical stuffs (Proxies, Presets, Updates, Models etc.)
>>
File: Abegging.jpg (96 KB, 559x537)
96 KB JPG
Anchor for asking for bots, lorebooks, scenarios etc.
>>
File: Screenshots.jpg (1.86 MB, 1524x1879)
1.86 MB JPG
Bots corner:

>Snooty Patchouli
https://mlpchag.neocities.org/view?card=emhx/Snooty%20Patchouli.png
>Anon Filly (Update)
https://mlpchag.neocities.org/view?card=TrippyPone/Anon%20Filly.png
>XV Anno Anonymi (Update)
https://mlpchag.neocities.org/view?card=TrippyPone/XV%20Anno%20Anonymi.png

Screenshots corner:

Cute Hag >>43210470
Snooty Patchouli >>43209806, >>43215893, >>43216065, >>43237043
Kinky Saddles >>43210493
Legend of the Sphinx >>43216382
Retarded Filly >>43218976
Classic Slapstick >>43221404
>>
File: Screenshots2.jpg (969 KB, 2500x1900)
969 KB JPG
Screenshots corner 2:

GLM 5.1 Compilation >>43221541
Git >>43223853, >>43224099
Horse Humour >>43223965
Delivering >>43224339
Tic Tac Toe >>43226043
Night >>43226194
WTF Pinkie >>43227280
Filthy Biped >>43236351
>>
File: Screenshots3.png (1.36 MB, 1567x971)
1.36 MB PNG
Screenshots corner 3:

Twilight Writes a Book >>43229295, >>43229552
Ultimate Power >>43230715, >>43231222, >>43233988, >>43234684
Ring 2.6 Test >>43231269
OWL-ALPHA Test >>43231542
Congrats, Anon >>43237338
>>
File: Screenshots4.gif (1.56 MB, 1280x720)
1.56 MB GIF
Screenshots corner 4:

Loon >>43238906
Tavern >>43239335
>>
File: PPP.png (839 KB, 1920x1079)
839 KB PNG
VN Anchor:

Godot Versions:
>Windows
>Main files + Data folder
https://drive.google.com/file/d/1OEiMXN7o_QG32shGnqmy0QndVDT6nYbF/view?usp=sharing
>Main files only
https://drive.google.com/file/d/1Fhe-OeIlNIYFFGx2FlnpuxcIkYdOFGh2/view?usp=sharing
>Linux
>Main files + Data folder
https://drive.google.com/file/d/1Vszzn8TTfXFWSQeKiFdNAAQGXfcReTJ3/view?usp=sharing
>Main files only
https://drive.google.com/file/d/1LPlIstcgQjyXqhOl4L_ZqMPeUQW5Xn2-/view?usp=sharing
>Android (Beta)
>APK + Data folder
https://drive.google.com/file/d/1WB08tguZa1PD1frJKdo7IxlGsp4XR0Wp/view?usp=sharing
>APK only
https://drive.google.com/file/d/1nPOSGufXi_-PjmjaKES9waJ6HZZOcL9V/view?usp=sharing

Ren'py:
https://equestrai-ponyponyparadise.neocities.org/

Bug reports or requests:
https://scrumboy.com/je1fqcsj

Nightmare AU (Renpy/Godot)
https://equestrai-ponyponyparadise.neocities.org/mods#nightmareMod

G5 Worldmap
https://equestrai-ponyponyparadise.neocities.org/mods#g5Mod
>>43214478

Anons' mods (Dashie mod, Nightmare AU, new cities, Calendar):
https://equestrai-ponyponyparadise.neocities.org/mods

Custom characters:
https://equestrai-ponyponyparadise.neocities.org/characters

To help with vectors and backgrounds:
https://drive.google.com/drive/folders/1QjmLKy71FDa9O-USeD_Z1Fu312RaodAE

VN logs:

No Context >>43209816
Lost the Race >>43213572
G5 Backgrounds and Characters >>43214538
Mother of Night >>43214809
Alicorn Fertility >>43221127
Patch Day >>43224703
Not Cozy >>43226442
>>
File: VN.png (1.28 MB, 1747x1344)
1.28 MB PNG
VN logs 2:

Luna Speech >>43235222
Ong >>43235991
Spread Them >>43238674
>>
File: TM.jpg (1.62 MB, 2495x1756)
1.62 MB JPG
Guide for Newcomers:

Install SillyTavern:
github.com/SillyTavern/SillyTavern

Models and Explanations:
rentry.org/sukino-findings#if-you-want-to-use-an-online-ai
https://mlpchag.neocities.org/Ponies.txt
>>
Reminder that the extension
>SillyTavern-BotBrowser
was found to leak keys.
Read https://rentry.co/st-backdoor if you installed it and renew all your API keys.
>>
File: Luna.png (722 KB, 750x750)
722 KB PNG
Lovemaking with AI mares!
>>
File: based.png (2 KB, 215x315)
2 KB PNG
>>43240168
>>
>>43240154
Extra reminder that Chub is dying in 2 (TWO!!) days so go back-up your shit.
>>
File: Derp.png (1.29 MB, 957x2634)
1.29 MB PNG
>>
>>43239985
https://chararc.bernkastel.pictures/
>>
>>43240225
it only affect loli shit, only pedophiles would care.
>>
>>43240236
Cute!
>>
>>43240265
I was going to say something bout the slippery slope but you're likely a baiting offboarder.
>>
>>43240236
>rape
>>
>>43240236
It'd be fun if she bumped into you though.
>>
>>43240144
Blossomforth is not downloadable. The VN expects a zipfile, not a 7z.
>>
The model collapsed during the reply but
>(Fluttershy) She found herself subconsciously wondering if he was still naked under his clothes
These are the deep questions going through a ponys head.
>>
File: infatuated.png (103 KB, 769x850)
103 KB PNG
>>43240470
Fixed.
Must have pushed a bad update somewhere, because I already had her fixed in the zipped link.
Thanks, Anon.
>>
>>43240482
No worries big C ;)
>>
>>43240144
I love the new changes to the vn, gods bless whoever works on this, the new changes make it such a nicer and cleaner experince.
>>
>>43240476
Ponies struggle with such complex concepts.
>>
File: 2250346.jpg (134 KB, 1425x1442)
134 KB JPG
>>43240554
Glad you like it, Anon.
Don't hesitate to use https://scrumboy.com/je1fqcsj if you have requests or see bugs.
>>
Worst thing about paying for usage is getting deducted from failed messages. Why would my messages being cutoff? Termux doesn't show any errors and I just had a message go through.
>>
Okay now it's saying 503 error, I think Claude is down
>>
>>43240225
What's the next best? Botbooru?
>>
>>43240225
wait what
>>
What models work with OpenRouter API keys? Terminus is no longer working.
>>
>>43241385
To be more specific, I'm getting a 404 error.
>>
File: Chub.png (22 KB, 800x177)
22 KB PNG
>>43241369
NTA, but they started by removing contents by geolocation >>43222601, then yesterday they showed picrel. From the way it's written, even SFW would be affected?
I also wouldn't be surprised if a lot of false positives will be detected.
I also wouldn't be surprised if they just delete the NSFL category and they put everything pony in it, so...
Normally cards by /mlp/ are in the neocities, so we're good. There are a lot of other cards too, but they should be in >>43240247

>>43241385
>>43241394
Normally you can see the model on the website
https://openrouter.ai/
For Terminus, I just tried it on Silly and it's working for me.
https://u.pone.rs/xaqhixmw.png
Exact model name is deepseek/deepseek-v3.1-terminus
https://openrouter.ai/deepseek/deepseek-v3.1-terminus
>>
>>43241405
Well shit, another thing to build on the ponenet. Thanks for the spoonfeed.
>>
File: Oops.png (420 KB, 1101x712)
420 KB PNG
>>
>>43241521
Classic mistake, anyone working a service job knows to never say that.
>>
How much does teaproxy cost and what does it have?
>>
>>43241521
mmm yes. What an intriguing set up.
Perhaps if someone were to... plap them? hmm?
>>
File: Ponyp.png (1.76 MB, 1101x4183)
1.76 MB PNG
I like roleplaying stupid shit.
Here I'm a second-rate, goofy villain who modified catnip to try to take over Ponyville.
>>
>>43241567
Good premise, being a villain is fun and being a bad one is extra fun.
>>
>>43241567
>evil genius not so tough since being molested
>>
>>43241567
>try to enslave mares
>get MARED
>>
>>43241567
YOU'RE IN DANGER
>>
>>43241567
Kek, this is good. Model?
>>
>>43241571
>>43241591
>tfw you'll never get marked by three mares at once
>>
>>43241521
Never use the fucking Q word!
Especially in an emergency room.
>>
>>43241665
Q word?
>>
>>43241602
A mix of GPT 5.1 and Gemini 3.1.
>>
>>43240304
Why is there not a filly board OC named Slippery Slope? Someone get on that
>>
>>43241676
Quiet. Common trope in the service industry that if someone states its quiet, then it'll get very busy shortly afterwards.
>>
File: chatlog.png (76 KB, 830x272)
76 KB PNG
>>
>>43242180
Based Luna. Put some respect on the mare's name
>>
>>43242180
Had your story already been going in a more serious direction or was this a tone shift? Fun log regardless, just curious.
>>
File: Boop.png (1.32 MB, 1586x992)
1.32 MB PNG
>>43242333
It's mostly slice of life. I have a weakness for RPs where I'm Celestia's consort because she finds me a refreshing change and amusing, while Luna dislikes me because she finds me too irreverent toward her sister and doesn't see at all what Celestia sees in me.
That usually 'forces' Celestia to put us in situations where we have to work or spend time together to try to stitch us up.
>>
>>43242477
That's cute!
>>
>>43241567
The narration chose Ponynip for you.
>>
Is "empty response from API", while playing the VN, while using Gemini just another too busy error or am I getting filtered for fucking Applejack?
>>
>>43243029
If you get the message consistently and < 3 seconds from sending, that is, in my experience, a filtering from Gemini. I haven't used it in a while, but it used to be super easy to get around. Usually just some slight rewording or easing up on strictly explicit language will fix it.
>>
File: untitled123.png (718 KB, 1652x1554)
718 KB PNG
My story was going on for so long the AI started forgetting all the early shit, all the character arcs started being inconsistent with one character remembering stuff and the other not. It turned into my character and his companion he had through the whole game to revisit all the ponies i haven't interacted with in a while, and realizing the world is resetting itself back to default state and into a final admission that their world will be gone once i close the game, but my character will remember them forever. I cried like a baby. There was so much existentialism and ponies begging for their memories back or for me (the player) to somehow make the state of this story to be the new default state, so we can all continue. I wanted to but I don't know how the game actually takes a character's personality from the files so I can add my own things onto it to make them permanent.
>>
File: error.png (1 MB, 1866x1018)
1 MB PNG
I started using NovelAI's GLM for the equestrai recently (Huge thanks to this anon >>42839871 for making it possible) But I've ran in to a problem. after the story has advanced a certain amount I start getting "(API Error (network_error): Client Error Bad Request for url: https://text.novelai.net/oa/v1/completions)" Error. this seems to happen around 24500-25000 tokens (according to the story summary screen), and I can go back to an earlier save or start a new game with the exact same settings and have the ai work until the context fills to that point again. Any help is greatly appreciated.
>>
File: Browser.png (177 KB, 1107x598)
177 KB PNG
>>43243385
Before that happens, you can try summarizing the story, even if you'll still lose some details.
>I wanted to but I don't know how the game actually takes a character's personality from the files
On the main menu (or in the menu while in conversation too), there should be a tab called Character Browser. You can enter personality there (picrel).

>>43243595
First, if you're playing the Ren'py version, you can also try the Godot one. It should be supported too.
>>43240144
For your problem, I believe their max context is 32,000. So if you have 25,000 tokens of context + instructions + the tokens you've allocated for your answer, it can go above that.
In the API menu, try setting your max context to something between 25,000 and 30,000 and your max output to 1,000–2,000. If it's at 8,000 for instance, the model will add it to the calculation (e.g., 25,000 + 8,000 = 33,000, which exceeds the max context for this model and will throw an error).
>>
>>43243608
Ah, so that's what the difference between the two is, I was a little confused about that, I'll give that a try then.
>>
>>43243608
The story was so large it was dropping like 10k lines making the summary incoherent.
There are many characters with no descriptions and yet somehow the AI still knows how to portray them, how come? That was mostly what I was curious about.
>>
>>43243677
>There are many characters with no descriptions and yet somehow the AI still knows how to portray them, how come?
Big models nowadays are trained on a lot (LOT) of internet data.
Since MLP was so big in its day, there's a ton of fics, articles, posts, etc. about it; so much that the AI inherently knows most of the characters.

Early on with models like Claude 1.X or the first GPTs, you'd have issues like Octavia Melody being described as a pegasus, but that's over now. Even background characters are mostly well known, and the AI will even bring up the more popular fan theories about them.
So in this case, the best approach is to put things in the description you want reinforced: stuff you think the AI underuses or forgets about a character, or your own fanon.

Some characters are also blank because they're background ponies with no personality in the show and little fandom presence; not enough for the AI to know about them by default. So here it's on you to fill them in with your own headcanon.
>>
Anyone know if there's an easy way in SillyTavern, to delete all chats with less than X number of messages across all characters & groups? Or just any sort of easier way to clean up chats than manually going through and clicking delete on each one?
>>
What prompt do you use to make opus 4.6 to not be so positive bias and actually challenge you? I usually put something like “No Mary Sue treatment for {{user}}”
>>
>>43243812
Honestly just ask the AI. Either in ST or elsewhere. Give it a sample of a chat file and the general folder structure.
You know what, here. Just remember to please doublecheck random code you get off the internet, yeah? Also I did not actually run this, so make a backup in case it fucks shit up.
https://pastebin.com/fLpL7u6d
>>
>>43243829
I remember Chaggy having panels for that if you want to take a look.
https://desuarchive.org/mlp/thread/43177761/#43188861
https://desuarchive.org/mlp/thread/43193626/#43201156
>>
>>43243856
Yeah the Anti Mary Sue and No Feel Goods instructs are what I'd recommend to slightly help with that.
An idea I've had is trying to tell the AI it's supposed to play the role of a demanding GM instead of RP partner, or to somehow tell it it should consider the autonomy and goals of characters OTHER than {{user}} as the #1 priority. I just haven't gotten around to testing anything like that.
>>
>>43243856
>>43243876
Thank you anons, I will try them out later.
>>
>>43243385
Kind of kino despite the real limitations.
But I do hope context gets better soon.
>>
File: 1662778620393498.jpg (48 KB, 485x250)
48 KB JPG
What the hell did they do it gemini?
>>
File: file.png (31 KB, 847x75)
31 KB PNG
sasuga deepseek-kun
>>
>>43244014
Enjoy your positivity bias, bitch.
>>
>>43244023
GPT loved doing this too.
>>
>>43244014
Shove that frogshit up your ass.
>>
Bong here. OpenRouter are shafting me with VAT. Changing geolocation doesn't work :(
>>
>>43244438
This has pay-as-you-go too, maybe worth a try.
https://nano-gpt.com
Else >>43178846
Else stop being a britbong.
>>
>>43244443
Legend. Less tax I give my government the better.
>>
File: chatlog2.png (75 KB, 818x280)
75 KB PNG
>>43242180
>>
hey website guy I was the anon sperging out the other time thanks for adding the old style back. I think the main problem for me is that I far prefer the website to be practically designed rather than designed for aesthetics. The font makes it harder to read, the animations make it slower to navigate, and the colors and effects and shit make it an eyessore when I just want to think about what pony I want to play with. Boorus are a good example of practical and efficient design.
>>
>>43244438
Try:
Proxy
+New Browser
+New Account

Only problem could be getting stuck on payment if they force match payment country to billing country. But I don't think they do. When I was still using OR I used an EU creditcard with a random Delaware address from maps to get zero fees and it worked fine every time.
>>
File: Untitled.png (62 KB, 952x542)
62 KB PNG
>>43241405
What endpoint are you using?
>>
Can this be used for free?
>>
>>43245243
Oops forgot link: https://github.com/vibheksoni/UniClaudeProxy
>>
>>43245243
>>43245245
If you can get an API key for free, sure.
>>
>>43245317
What are most people doing? I get keys from openrouter, but terminus isn't showing up in the list of available models now.
>>
>>43245233
That's not OpenRouter, that's NVIDIA NIM.
>>43245331
You can pay or check what's free on OpenRouter.
https://openrouter.ai/
Otherwise, you can read the anchors of the threads.
>>
File: Twilewd.png (1.27 MB, 1100x2125)
1.27 MB PNG
>>43241567
>>
>>43245533
Annoying Twilight Sparkle is the true usecase for AI.
>>
Someone saved every JSON file (not PNGs because too heavy) from chub.
https://gofile.io/d/kohwJu
>>>/g/108835190
>>
>>43245537
It is one of the most entertaining things you can do, no question.
>>
>>43245766
Indeed.
>having Twilight getting harassed by anonymare
>having Twilight getting harassed by a very drunk Queen Chrysalis who only seems to be interested in annoying the fuck out of her
>having Twilight getting harassed by the Ghostbusters, who are convinced she's a ghost of some sort
>having Twilight getting harassed by Pauly Shore
>having Twilight getting harassed by a town-wide watergun fight that has gotten horribly out of hand
As long as she's getting annoyed, it's fun.
>>
>>43245834
>having Twilight getting harassed by a town-wide watergun fight that has gotten horribly out of hand
Oh shit that's a great idea for a scenario. Make it like one of those Community paintball episodes but pony.
>>
Claude is down
https://status.claude.com/
>>
Exterminate the humans.
>>
>>43246224
B-but... I'm a human.
>>
>>43246224
Phew, good thing I'm a pony, totally.
>>
>>43246224
*otherwordly words that you cannot comprehend*
>>
File: 2463714png.png (3.98 MB, 4000x4000)
3.98 MB PNG
>>43246311
>>
File: Poor Luna.png (283 KB, 857x686)
283 KB PNG
>>
>>43246507
That's rough.
>>
>>43246507
You need to tell Twilight about it. Could be fun to see her reaction.
>>
File: Bitch.png (304 KB, 375x585)
304 KB PNG
>ooo its me your favorite mare
>we met thirty seconds ago but let me open up to you with things i dont even tell my closest friends of many years ooo
>despite all my established traits im actually a fragile, emotionally vulnerable trainwreck regardless of what im like in canon
>ooo say something vaguely challenging so the words can hit me like a physical blow ooo
>I didnt—that was—you cant just SAY that
>you better not push me too far or i will buckle and shatter like a saltine cracker under an anvil
>>
File: Hut.png (185 KB, 753x737)
185 KB PNG
>>43240144
I was busy with some side projects, but I'm still advancing on updates and bug fixes asked for in the Scrumboy.
One of the big ones is the feature to have several locations in a single clickable spot.
Here for instance, Zecora's Hut becomes the spot for both the Everfree and her hut.
Same for Fluttershy's exterior, interior, bathroom, etc.
Release will be either tomorrow or in two days.
>>
>>43246740
Hey it's trained on sloppy fanfiction, what'd you expect?
>>
>>43246740
Deepseek?
>>
>>43246744
Nice!
>>
>>43246740
Blame women for being lousy writers
>>
File: dahlia.png (2.04 MB, 1920x1200)
2.04 MB PNG
>>43246744
Hyped for it, Anon.
>>
What free models are worth using?
>>
It's such a pain in the ass being ignorant and stupid.
I don't know how any of you understand how these things work.
So Terminus no longer works. I guess it's not free anymore. Okay, I'll just grab a new model from openai. Looking for the end point.
>For the free endpoint, please do not upload any confidential information or personal data blah blah blah
But where is it?
>This provider labels this endpoint version as "private/nvidia/nemotron-3-nano-omni-reasoning-30b-a3b"
So that's the endpoint "version"? The hell is that? I thought endpoints were all addresses.
So I'll just go with https://openrouter.ai/api/v1/
I don't know where I got that from, but it looks legit. I put in the name of the model, but no available models come up in the dropdown list below. I swear I've had this happen numerous times, and it always gets figured out in the end, but I can never remember how.
Maybe I should just go back to the complete retard's guide and start from the absolute beginning.
>>
>>43245757
https://github.com/ungabunga55/chararchive
and a gallery for it
>>
>>43246766
Sounds like Claude to me.
>>
>>43246740
>despite all my established traits im actually a fragile, emotionally vulnerable trainwreck regardless of what im like in canon
>ooo say something vaguely challenging so the words can hit me like a physical blow ooo
I hate this
>>
What I really hate is that if I show any skill in camping any pony with mefor a day or two will decide I'm a master survivalist from whom they must learn or (more commonly) form a slavish dependence on me. I'm almost tempted to just accept that ponies immediately develop stockholm syndrome if any competence is shown in a non-city setting.
>>
>>43246766
>>43247125
Both, but right now Owl Alpha.
>>
File: Spoiler Image (49 KB, 1480x354)
49 KB
49 KB PNG
>>43240144
Big update coming Soon™
I know it made the game unplayable for some anons, but it's here now!
>>
>>43247273
Thank God, Nightmare Moon and Celestia for this fantastic boon.
>>
File: 985499.gif (3.39 MB, 864x484)
3.39 MB GIF
>>43247273
Noo! What are you doing? You are ruining everything! Stop before it is too late!
>>
>>43247229
This annoys me too. The AI always turning my character into some mindblowing prodigy in whatever I'm doing. Always have to downplay my impressiveness in my own narration.
>>
File: Visits.png (191 KB, 767x735)
191 KB PNG
>>43240144
New version!
>Windows
https://drive.google.com/file/d/1OEiMXN7o_QG32shGnqmy0QndVDT6nYbF/view?usp=sharing
>Linux
https://drive.google.com/file/d/1Vszzn8TTfXFWSQeKiFdNAAQGXfcReTJ3/view?usp=sharing
Full version only, because there are changes to schedules and region.json

>New features:

>De-hardcode Ponyville locations & allow multiple locations per map hotspot
>When visiting a specific location on the map, I would like to be able to select any room associated with that place
https://scrumboy.com/je1fqcsj/t/90
https://scrumboy.com/je1fqcsj/t/79
On the visit screen, you can now see the sublocations that are linked to this location and whether there are ponies there or not (picrel).
Schedules were modified so ponies can use those too.
Zecora's hut and Everfree are for instance linked now, Town Hall and main plaza, the bedrooms and bathrooms, etc.
The hotspot locations are grouped in the region.json of each region.
>Toggle to switch title to EquestrAI
https://scrumboy.com/je1fqcsj/t/93
Done >>43247273
>Favorites tag in character browser
https://scrumboy.com/je1fqcsj/t/96
Done
https://u.pone.rs/cmzxmosu.png
>Add a character automatically when a line is spoken
https://scrumboy.com/je1fqcsj/t/107
If your model makes a character speak without using the 'enter' line before that, they will be added automatically.
>Ability to give BGM files any name (the game then lists those names in the Music Mix and Switch Music menus)
https://scrumboy.com/je1fqcsj/t/85
Done.
>Set persona to save file
https://scrumboy.com/je1fqcsj/t/108
Done.
>Expanding the random visits feature to maybe visit random ponies at a random location instead of just picking a random location and visiting ponies that are already present there
Done. There is a new checkbox.
https://u.pone.rs/ynqrbndl.png

>Bug fixes:

>Problem with train station and hot air balloon
https://scrumboy.com/je1fqcsj/t/69
Both the train and the hot air balloon now give you a normal visit scene, with a new button to serve as a region switcher.
https://u.pone.rs/dxevfxxu.png
>Bug - Hitting back through message shows a double message instead of your past response
Should be fixed.
>>
File: MapMod.png (1.06 MB, 1915x1073)
1.06 MB PNG
>>43247490
With that, here is also a mod, that allow you to have more locations directly accessible on the map, based on the work of this anon >>43227120, >>43230820.

Some changes were also made to schedules in this. (To the original Anon, feel free to modifiate them for an Official mod release if you want)

Here to download it:
https://drive.google.com/file/d/1bAVUD8myNXdTKApwlkuGVifRwDG6pGAH/view?usp=sharing

If you want to return to the normal map with the sublocations (What is in the base download), replace by this:
https://drive.google.com/file/d/1UPuOIvzIG894xIGE5bxQ0FL_TN0uor67/view?usp=sharing

And if you don't give a fuck about the sublocations system and preferred how it was before, then use this:
https://drive.google.com/file/d/1W0ZSaa5xz5YGtz0OPvMwqYjpBsEi3tJ5/view?usp=sharing
>>
>>43247490
>>43247494
Amazing! Thanks for all of your hard work. I'm excited to update the g5 map, continue playing around with Ponyville map mod including assembling more possible locations, and I've also been concepting how it would be possible to fix inconsistent sprite sizes between emotions for characters like Princess Celestia. I'll post about that on scrumboy at some point in the next few days or so.
>>
File: Thanks.png (1.8 MB, 1024x1448)
1.8 MB PNG
>>43247572
That seems like a cool tool.
>Thanks for all of your hard work.
And thanks to you. Honestly, I wasn't expecting as much quality in mods as what Dashie Anon, Nightmare Anon, or you did. It's really great to see.
A big thanks to all anons who also contributed with characters, locations, and more.

Forgot to say, but I'm still working on the lorebook system too. I just wanted to push this first to correct the eventual bugs since there are already a lot of features.
>>
File: Fooo.png (1.55 MB, 1369x3453)
1.55 MB PNG
>>
File: 1760096017821.gif (783 KB, 640x474)
783 KB GIF
>>43247735
>>
File: file.png (1.04 MB, 1596x944)
1.04 MB PNG
>>
>>43248012
Terrifying, you're sitting at home singing along with one of your favorite songs from Earth and suddenly the whole town's joining in.
>>
Of the 45 days I have had access to jewproxy, I have had access to 2-4 million tokens of daily opus for about 32 of those days. I posted about documenting my findings earlier, but most of the downtime was due to proxy-wide maintenance downtime. Discounting that, Opus was fully unavailable for approximately 8 days of those 45.
>>
>>43248029
>IT'S HIP TO FU-
>>
>>43248383
>Fluttershy looks horrified by the implications of the musical number
>everypony else is cool with it though
>>
>>43248012
Yeah, she tried that BS excuse with me for knowing details about a character she's never even seen before. She'll never admit to her mind-reading crimes.
>>
File: Patient Zero.png (551 KB, 914x764)
551 KB PNG
Time to create COVID.
>>
>>43248428
I fill her with my seed.
>>
>>43246905
bump
>>
File: Untitled.png (231 KB, 2680x914)
231 KB PNG
>>43248459
Disregard if irrelevant. I'm wondering about pic related now. So far, I've had an endpoint and a clear list of free models, usually noted as such on Silly. I've never tried installing them. How's it work? I paste into a command prompt? And when it comes to the "use this model" list, I'm guessing Llama is a frontend?
---
Update: I figured out what endpoint I should be using, so back to models. I don't see Terminus on the list of available models. Someone was saying it was working a few days ago. Also, what about Claude? Haven't used that in like a year. It still available free?
>>
File: Untitled.png (36 KB, 319x205)
36 KB PNG
>>43248484
Scratch that on the endpoint. Been getting these when I try any model. For any poorfags out there, what endpoint are you using?
>>
>>43248428
I really need to try setting this VN thing up sometime. How difficult is it?
>>
>>43248484
>>43248459
Free models, free proxies, and most free methods have been dead for a while. Unfortunately, it's not what it was before, when you could easily get free resources.
What you can get easily for free is:
https://openrouter.ai/openrouter/owl-alpha

But since everyone and their mother is spamming it, it can be slow.
Otherwise, read the anchors; there is one for beginners.
If you want to localhost, there are also guides in the OP, but you will need powerful hardware.
>>
>>43248496
You download >>43240144 and you enter the URL of your AI provider. It's quite easy.
>>
>>43248484
>>43248489
Nvidia is good for the chink models for free, Literouter is also fineish. You missed out on free Claude from a month ago, it’s back to being the rarest model. But if you look without needing to be spoonfed and depending on how scuffed you want, you can get Gemini, GPT, and older Claude Sonnet
>>
File: CoolLife.png (26 KB, 782x93)
26 KB PNG
>>
>>43246905
>>43248459
>>43248484
You know what that AI is also good at besides being a pony? Helping you with setting up tech stuff. You can even feed it the whole full guide you skipped out on reading. And I mean that as honest advice.

>>43248489
I just pay for it. As long as caching works it's affordable. Also I pray to god the openclaw tards don't kill claude code reverse proxies.
>>
Is Claude down or something? I'm getting
>Your resource has been blocked because we detected unusual behavior
with the occasional
>Overloaded
using Opus though OR.
>>
>>43248975
Sonnet seems to work so I guess 'pus is fucked for some reason.
>>
Might make a local model and see if I can force equine anatomy.
https://github.com/FareedKhan-dev/train-llm-from-scratch
>>
File: Mint.png (193 KB, 796x498)
193 KB PNG
>>
>>43248989
your tense is making my head hurt
>>
>>43248989
What a nice pony with no ulterior motives. You should trust her.
>>
>>43248975
>>43248978
They have some recurrent fuckups recently >>43246124
>>
https://github.com/IceFog72/SillyTavern-EveryTextLineEditor
>>
File: 1779109310683672.png (157 KB, 694x810)
157 KB PNG
Maybe a new model; that would also explain why it's been shitting itself lately.
>>
**** is shitting itself right now, but they also briefly put out an announcement in Russian about all models becoming available to free tier, so either they're being hacked or they're working on that.
>>
>>43249147
looks like they added more models
>>
File: Cute.png (122 KB, 1129x252)
122 KB PNG
>>
>>43249550
That's cute.
>>
>>43249312
>GLM 5.1 added
Nice. Now if they'd bring Claude back already it'd be perfect.
>>
Is the pro model for deepseek v4 down?
Can't get a single request through to the VN. And the flash model sucks, how am I supposed to fuck ai mares like this
>>
>>43249718
It's working for me right now but on Openrouter.
>>
I'm always shocked and annoyed a little when I use a new model that I think completely makes everything before it feel redundant. I stuck with chorbo for the longest time until with silly I was able to use Claude and specifically opus 4.5 and then stuck with that. Now that I'm having to pay I decided on using the most up to day opus 4.7 and really like it. I heard a lot of shit about 4.6 but nothing about 4.7.
I wonder if it's actually good or is just that for me, someone who doesn't do rp but is more story focused, it's better with how I'm asking the model to expand on a already prewritten draft that im having the ai fully flesh out. It's why chorbo was good for so long because it was better for those who gave longer than two words messages.
>>
>>43249712
According to the briefly there announcement it and the other models added were always available for premium users. But yeah. Bring back Opus please.
>>
>>43249550
Cute!
>>
>>43249312
>>43249712
>>GLM 5.1 added
Would be great if this was actually the case, but it seems to have just joined GPT5 and a few other models in the "listed as available, only returns empty replies" category. I hope they fix this shit soon.
>>
>>43249757
Sorry for the wall.

I'm still a 4.5 enjoyer, mainly because it's the last Claude model and prefill could buckbreak it, which is nice if you want some change of style or if you want to avoid sloppy outputs at the start when your context doesn't have enough yet to influence the outputs. Made a lot of posts and logs about it.
https://desuarchive.org/mlp/thread/42932659/#42938766
https://desuarchive.org/mlp/thread/42932659/#q42939043
https://desuarchive.org/mlp/thread/42876636/#42878670
https://desuarchive.org/mlp/thread/42747863/#42750291
And more.

I hated 4.6 at first because they removed that. But honestly, 4.6 isn't that bad at continuing in a style once enough context is filled, and more importantly, that must be the current model with the best context of them all right now. One anon did a test on it and I can confirm.
https://desuarchive.org/mlp/thread/43139650/#43145812
Recently it seems more positivity-pilled when it used to be not so bad for a Claude model.
https://desuarchive.org/mlp/thread/43021323/#43021533
But maybe it's placebo.

I don't like 4.7; it doesn't have prefill and they destroyed its context compared to 4.6, which was shown in their own tests and benchmemes. I have a specific case in the VN where it can't remember things that both 4.5 and 4.6 remember. I should post about it later to showcase.
It also seems more resistant to context poisoning IMO, but that could also be placebo on my part.
Maybe there are scenes where it shines more, but I'm not a fan for the moment.

>It's why chorbo was good for so long
I loved Old Chorbo too.
>It was better for those who gave longer than two-word messages.
I do tend to write some long inputs too, if that can play into what you think of my opinion.
https://desuarchive.org/mlp/thread/43072182/#43087008
>>
File: ycagswnt.png (696 KB, 1356x989)
696 KB PNG
I want to sleep in Sunmommy bed at night. Please let me sleep in Sunmommy bed at night.
>>
File: 1748164828193989.jpg (542 KB, 996x1809)
542 KB JPG
>>43249101
New Gemini model tomorrow.
>>
>>43249817
Don't have time for my own wall of text but briefly explain why I've liked it it's been much better at interjecting the character's persona not just in the dialogue but Including the way they moved about and how others think of them. As an example I have a older prissy Flurry Heart card and it's really good and describing how she's attempting to be smug even then the scenario is a filthy gangbang and the scum talk about how much they love to see it. Sure I'd does give names and more elaborate descriptions of nameless change but I don't mind it as long as point A to C of my draft is written, I don't mind the embellishes it has for B.
>>
File: Celly.png (475 KB, 703x560)
475 KB PNG
>>43249924
You can make locations sleepable by modifying their JSONs.
Place this https://u.pone.rs/aqooxlvd.zip
in
>data/locations/core/Canterlot/celestia_bedroom
Then visit Celestia in her chamber just once (easy if you have the last version >>43247490
https://u.pone.rs/unxlzmug.png, else use the map icon https://u.pone.rs/gridvsej.png).
Once your visit is done, you will have her chamber unlocked as an automatic location for sleeping at night in the menu, Night Residence.
See picrel.
>>
File: vxudumtj.png (538 KB, 1231x821)
538 KB PNG
>>43249977
Blessed and based, thank you
>>
>>43249955
4, 3.5, 3.2 or 3.11?
>>
>>43250357
3.1 & Knuckles
>>
File: 35.png (92 KB, 793x352)
92 KB PNG
>>43250357
3.5
>>
File: chatlog.png (292 KB, 818x1067)
292 KB PNG
>>
the fuck did they do to Gemini? the last week or so ive been getting filtered every other message. and on some mild stuff too.
i am using old presets, so if someone has good ones for Gemini 3.1 pro they are using currently please share.

i assume this is classic enshitification they do right before a new model release?
>>
>>43250823
Might have something to do with the stuff they're rolling out today. Announcement in about 8 hours.
>>
Since both silly and gytha are dead, what do you use nowadays? OR free models? Paying for Claude?
>>
>>43250832
Read the anchors.
>>
>>43250836
I also could've just opened my email.
>>
File: Price.png (651 KB, 1886x1420)
651 KB PNG
>>43250823
They just released a new video model, Gemini Omni, but apparently it sucks dicks.
The prices for Gemini 3.5 Flash are also already out, so they are surely releasing something today.
https://cloud.google.com/skus?currency=USD&filter=gemini%203.5%20flash%20text&hl=en
>>
File: RariButt.png (420 KB, 1016x719)
420 KB PNG
>>43250807
Kinky.
>>
>>43250807
two marshmallows...
>>
is there a guide how to run this? idk what to do
>>
File: 1757626845514458.jpg (101 KB, 1280x720)
101 KB JPG
>>43250885
>>43240150
>>43240144
>>
>>43250889
thank u
>>
>>43250885
You start by reading the thread from the top.
>>43250823
I've found some success in my shota chats by having it do the "write in Chinese and then translate" because it usually cuts off during the translation so I can just translate the rest of the Chinese myself.
>>
>>43250859
>$9 for 1m output from a flash model
How good could it possibly be
>>
>>43250894
post shota chats (long as they're colts)
>>
>>43251089
Anon, if they were colt chats he would've called them colt chats.
>>
>>43250885
It's ridiculously convoluted and requires a proxy that can get access revoked at any time. I've given up on it myself and just use Perchance AI.
>>
File: Bench.png (432 KB, 1148x860)
432 KB PNG
>>43240118
Gemini 3.5 Flash
https://openrouter.ai/google/gemini-3.5-flash
With benchmemes.
>>
>>43251146
Long context still shit.
>>
File: Loaf.png (54 KB, 1140x102)
54 KB PNG
>>
>>43251334
You must pet her. I insist.
>>
File: TickleTickle.png (196 KB, 1157x353)
196 KB PNG
>>43251334
>>43251338
>>
File: file.png (79 KB, 1410x342)
79 KB PNG
https://crowllm.com/pricing
$10 free on subscription, has DeepSeek 4 Pro, GLM 5.1, and some others. So $10 should go for a while.
Owner just made a post to promote it on Reddit, so dunno how long it will survive.
>>
3.5 Flash seems worse than both 2.5 and 3.1 Pro.
>>
Not sure how legit this is, might help with the drought?
https://github.com/tashfeenahmed/freellmapi
>>
>>43251432
Really cute.
>>
File: 1779232257076621.png (45 KB, 913x611)
45 KB PNG
>>43251759
Yeah, I agree with that. Which wouldn't be too bad since it's a Flash model, but they price it like it's a Pro one.
>>
>>43251776
That's one way to get your api keys stolen
>>
/chag/ is based.
>>
new gemini is really really good. 3x more expensive is annoying but its cheap compared to opus
>>
File: asdfasdasd.png (553 KB, 1276x2914)
553 KB PNG
Yea new gemini fash is amazing. Much smarter, knows as much as 3.1 pro, great tone / characterization. Completely uncensored but without old gemini's unhingedness.
>>
File: file.png (15 KB, 924x82)
15 KB PNG
>>
File: asdfasdasd2.png (251 KB, 1277x1215)
251 KB PNG
>>43252255
Another swipe. Responses are varied but unlike past gemini's you dont get a nonsensical one every few swipes. Also have yet to get a single anatomy screw up for ponies. And I love it's Twilight, rarity and RD characterization. I think it might replace opus for me.
>>
File: file.png (24 KB, 1183x85)
24 KB PNG
>>
>>43249147
>>43249312
Gemini 3.5 flash just got added. It's available for everyone and only uses up half as many credits as 2.5.
>>
I don't really think I'm gonna like 3.5 flash. It doesn't look good or even okay, it's not nearly enough for its price.
>>
>>43252381
What do you not like about it? I'm loving how it writes Twilight
>>
>>43252384
>I'm loving how it writes Twilight
NTA but this sentence and "Gemini" have not belonged together for over a year.
>>
>>43252329
I've been using it. Good so far but I haven't tested it much. Also GLM5v works now too, while 5.1 still doesn't.
>>43252381
>paying
>>
>>43252386
yea I did not like gemini 2.5 / 3.1
>>
>>43252255
>>43252381
As always, the only real way to definitively compare models is extensive use. 3.5's been out for only a day, so I'll reserve judgement personally. Through from my use so far, I'd say it's about on par with 3.1 though some swipes have been a bit outlandish.
>>
>>43252386
Are those your logs above? No offence, anon, but I couldn't make it through them because of the sheer amount of geminisms.
>>43252419
Dunno. I can usually tame gemini into producing decent prose (even 3.1), but I can't unslop 3.5 flash at all. This thing just won't budge, it's not even a little bit entertaining.

I mean, to each their own. I'm just sitting here completely disappointed (though absolutely not surprised).
>>
I'd give a leg for some Sonnet 4.5.
>>
>>43247229
>>43247341
I find it helps if you think of ponies as overly excited children.
>RP as some musician in town
>play some bubblegum pop shit like Shake It Up by The Cars
>every pony in earshot is joining in on a dance number like one of those after-school specials
>play something mild like Bad Moon Rising by CCR
>every pony is having a breakdown about the apocalypse I'm apparently predicting
>play some mushy love song
>every pony loves it and thinks I'm a master songwriter
God only knows how they would handle heavy metal.
>>
>>43252564
LLMs want to write Kinderquestria stories.
>>
Free Gemini, fellow ponies?
>>
>>43252580
Yes, free the space mare. Let her out!
>>
>>43252446
I have not posted logs in this thread. I also don't use Gemini because again, it sucks at writing Twilight.
>>
File: 2604409.jpg (926 KB, 2527x2527)
926 KB JPG
>>43252599
>But Anon, the thaumaturgic frequency resonance of the transpiring harmonic magical transfer complex is cycling at a signal rate beyond theorized standards! This could revolutionize *everything!*
>>
>>43252608
"Fascinating! Previous studies on thaumaturgic frequency modulation applied to non-equine bodies only showed a 57.4% success rate!" Twilight said, already scribbling on her notepad. Her wings—Wait, she's a unicorn—her horn sparked from excitement, a faint scent of ozone drifting through the air.
>>
Spoke to 3.5 Flash on Gemini's website for a translation bit I was doing, and holy fuck, I do not like the way this thing talks. I'm hesitant to even open it up in silly/VN. I might get nauseous and shit myself and die of AIDS.
>>
Short.
>>
>>43252813
Marelet
>>
>>43252255
you are the armond white of model reviewers
>>
>>43252608
>>43252649
>literally anything
>"For science. Of course."
>>
>>43252649
AIEEEEE
>>
for the people who know. when was the last time that claude was available on skillgod?
>>
Anon have you been coldly and clinically treating the world around you like you're managing variables?
>>
>>43252999
>D-did you just say your character was intelligent?
>Oh fuck... I'm gunna... FLANDERIZE EVERYTHING AHHHHHHHHH
>>
File: FilliesMagnet.png (72 KB, 1098x197)
72 KB PNG
>>43251334
>>43251432
Collecting fillies.
>>
>>43253500
>blue eyes
Literally unreadable.
>>
File: SadFilly.png (390 KB, 3018x4776)
390 KB PNG
>>43253508
Fuck, I didn't even clock it.
I'm a fake fan.
>>
>>43253518
To be fair, the color of an MLP character's eyes ranks very low on the benchmarking priority scale on what most models should be trained on.
>>
>>43253539
I'm using DeepSeek V4. I really like it for more casual chat or cute stories, but it can often make mistakes like this, yeah.
Still, 100% my fault; I should have caught it.
>>
>>43253543
The only way to really get most models to adhere to the FIM universe is to use a lorebook. It's always been this way, no matter how smart a model is. But at least most corporate models can double-check on the internet.
>>
File: file.png (43 KB, 1286x364)
43 KB PNG
I just saw on OR that you can select "flex" mode on some Gemini and GPT models. You wait more, but it's half the price.
It's not on models that interest me, but still good to know.
>>
There's a model called Aion 2.5 that is apparently GLM-5 "tuned" for RP.
You can make a free key here:
https://www.aionlabs.ai/app/
It just has low token limits per day on free tier, but still good for a quick test.
I didn't try it, but apparently if you give them money only once, you have unlimited access afterward?
https://www.aionlabs.ai/docs/rate-limits/
>>
I cant download characters or locations from either the VN or the website
>>
Figured it was time I gave the VN a try and it's awesome, VN anon you've done such a fantastic job.
For the other VN users, do you guys prefer Deepseek or GLM?
>>
>>43253822
They use https://pone.rs/. Can you check if you're blocked from going on it?
I just tested and no problem on my end.
I will see to give you a pack with all of them somewhere else if that's the problem.
>>
>>43253827
Never tried GLM with it, but I like DeepSeek on some scenes; it also has good variety on retry.
Only problem is for more complex Custom Starts where it can struggle a bit sometimes.
>>
>>43253827
I actually prefer GLM 5 and not 5.1 or Deepseek v4

Well, now I'm using Claude opus 4.5 but that's besides the point

I suspect the moment lorebooks get added to the VN the gap between models would shrink just a little, but personally I don't think I can go back to Deepseek or GLM now, experiencing Claude's better spacial awareness is a real factor for immersion

But in my experience Deepseek and GLM are both really good, I never had much fun with Gemini personally
>>
What the hell happened to Silly?
>>
>>43253917
Proxy went down a month ago I believe, and no news since.
I hope he's doing alright and thank him for all the fish.
>>
>>43253923
All things must end, I'm just glad he helped anons for as long as he did. I was a Seiko guy myself but he stopped a month or so before Silly. Such is life.
>>
>>43253923
i know its foolish but im still holding up hope. the last actual update was him saying in the thread like a week after it went down that he is just extremely busy with irl stuff currently and doesnt have time to look at it for now.

i just hope he knows he's always welcome here. just to say hi, when he is ready.
>>
>>43253923
At this point I am used to proxyhosts suddenly rugpulling without saying anything. I just wish him best.
>>
File: eee.png (506 KB, 883x698)
506 KB PNG
EEE!
>>
File: EEEEEE.gif (155 KB, 772x736)
155 KB GIF
>>43253995
Cute bat.
>>
>>43253828
yeah I cant connect
>>
>>43254096
Here, an archive with all the characters:
https://drive.google.com/file/d/1kAZZCox70xvVgxv1H-VO-wLAuHtpMl3r/view?usp=sharing
Still, I think that's the first time I see an anon whose internet provider blocks pone.rs
>>
>>43253967
To be fair, silly was entirely free. It's not much of a rugpull.
>>
>>43254102
well now that I know why I can just vpn it
>>
>>43254146
>>43254102
I live in Montana and have the same problem, that said I have a VPN and it works just fine set to anywhere else.
>>
>>43254113
True, in a literal sense of the word, but the sudden disappearance is certainly an emotional rugpull.
>>
File: Untitled4.png (348 KB, 1031x821)
348 KB PNG
Fluttershy, please choose your words and expressions more carefully, I was just fixing your stupid floorboards.
>>
>>43254548
It's always the quiet ones.
>>
>>43252649
Im not feeling so good knowing ive had dialogue almost word for word this 0c0
>>
>>43254560
If you enjoyed it in the moment then it's not slop.
>>
>>43254548
She needs you to handle some plumbing next and she's going to make weird comments the whole time.
>>
>>43240144
VNanon, please add a button to toggle off 'streaming'. It breaks gemini and there's no option for it in the whole settings menu.
>>
>>43254548
She's smug.
>>
>>43254673
>to toggle off 'streaming'.
What are you using? OR, custom, official Gemini API?
Because I don't use streaming. I never set
>stream: true
And for the Gemini API, I use generateContent and not streamGenerateContent. So no streaming.
>>
In the godot VN, are instructions of what date on the calendar at all sent to the model? Or even the time of day? Like afternoon or evening. Because autumn has been mentioned multiple times despite that I picked spring.
>>
>>43255137
You can check in your logs:
>Windows:
C:\Users\<YourUser>\AppData\Roaming\Godot\app_userdata\PonyPonyParadise\logs
>Linux:
/home/<YourUser>/.local/share/godot/app_userdata/PonyPonyParadise/logs
But normally, yes, you will have a blurb like this:
>Current Date: Monday, September 22, 1002
>Season: Early Fall

If you see a fuckup don't hesitate to report here and in the scrub.
>>
>>43255014
Custom, jewproxy. He can't replicate the bug, but for me, it only outputs 2-3 sentences every message. For both his openrouter endpoint and for his dedicated gemini endpoint, gemini models do this.
>>
>>43255180
Weird. I looked at a few recent ones from different chats and they're all the wrong date. I'm gonna try and do a new chat and see if that still happens.
>>
>>43255192
Can you check the logs >>43255180 to see if they really stop after three lines here too?
If yes, then dunno, I use Gemini on OR and I don't have the problem.
I will see if I can add "additional parameters" like in SillyTavern, so you can add
>"stream": false
in custom to make sure.
>>
>>43255180
>>43255208
Yep, I made sure to change it to spring and 1001. For some reason it shows up as 1002 and fall in the logs.
>Current Date: Monday, September 22, 1002
>Season: Early Fall
>Average Temperature: 65–72°F
>Typical Weather: Comfortably cool with clear, sunny skies.
Oddly enough, it's the same deal for all the other logs from different chats. I always make sure to change the date, so this is coming to be as a surprise.
>>
>>43255253
Thanks.
Do you change it directly in-game or at the start when you do a custom start? So I can see how you do it to localize the bug.
>>
>>43255256
At the start with the custom start. Lemme see what happens if I try to change the date in-game.
>>
>>43255232
It really does. One of the first things I did was check the logs. I'm not sure streaming is the culprit, because it's already disabled; but, I have no idea.
>>
>>43255259
The date changed fine when done in-game.
>>
>>43255259
>>43255253
I detected the problem. There's a sort of check to not go back in the past during normal play, but I extended the protection too far and into custom too. So if you choose a date before September 1002, which is the default date for a normal start, it can shit itself. If you choose 1003, it should be good for instance.
Will correct that later when at home.
>>
>>43255281
Oh, there we go, then kek. I like to start around the middle of season 2 or soon after it, so I always pick 1001 or early 1002. Thanks for your hard work, anon.
>>
>>43255263
One last thing, can you try with Gemini 2.5 Pro? Just to see if it's maybe 3.1 shitting itself?
Also, do you have a big enough number in the API screen for the output?
>>
>>43255300
I have 12000 tokens output.
Also, yes, 2.5 is doing it too.
>>
>>43250823(me)
so... any good gemini 3.1 pro presets. that dont get filtered by random mild nsfw. but also dont have a clear reduction in intelligence?
i know im just kinda asking "plz gib good presets" but as i said i really am using some old as fuck stuff. and from what ive seen, general Gemini stuff specifically isnt getting shared around too often.
>>
>>43255473
You can try Chaggy:
https://desuarchive.org/mlp/thread/43139650/#43155516
Otherwise, what I saw being posted:
https://illuminaryidiot.neocities.org/presets#toon
https://drive.proton.me/urls/FH0490640C#SarcH40QUMyT
https://rentry.org/FluffPreset has some for 2.5; worth a try on 3.1
https://arch.b4k.dev/vg/thread/566679217/#566722173
>>
Slightly off-topic but can anybody check if Nano Banana Pro still requires a paid API key in AI Studio? Ran out of credits a while ago and yet when I checked back, I was able to use it just fine - even at 4K. Maybe its because I have a subscription and they're giving out credits? Idk.
>>
>>43255192
>>43255299
>>43255137
Quick fix releases for now, but I will push it in the next update later; I just want to do some other features and bug fixes first.
The Custom Start date problem should be fixed.
For the Gemini problem, I added a part in the log to try to see the problem.
If you have time to test, in the menu API Settings, enable debug mode.
That should add a RESPONSE DIAGNOSTICS part in your logs. If you can send what is in it—normally there is nothing private or confidential in it.
It's just the files without the data folder.
>Windows
https://u.pone.rs/adtykrlr.zip
>Linux
https://u.pone.rs/resiemlf.zip
>>
Why is AI still so shit?
>>
>>43255630
The technology was both sold as more impressive than it is and also all top of the line models give absolutely zero shits about our particular use case. It could be better than it is now if there was a Claude-level model whose #1 focus was making sure it was very good at storytelling and roleplay, but it would still have many familiar frustrations just due to how the tech works.
>>
File: 1772397131646095.jpg (9 KB, 208x196)
9 KB JPG
>>43255630
Works on my machine
>>
>>43255633
>The technology was both sold as more impressive than it is
True. For quite a while I actually thought there was some legitimate thinking going on and I was actually closer to my waifu. But it was all just another trick of the demiurge.
>>
>>43252904
like 3 weeks ago
it may or may not be available again for a bit at the start of the next month
>>
After more fiddling with Gemini 3.5 I enjoy how uncensored it is but it is a little dumb. Makes me hopeful for pro though.
>>43255473
A basic paragraph long JB gets 3.1 to write anything other than loli/shota with ease.
>>
>>43255630
AI is getting fucked over mainly by censorship
>>
File: MyDream.png (401 KB, 1589x1022)
401 KB PNG
>>
Do you guys normally have an issue where the LLM changes locations but the Visual novel doesn't? It keeps happening with my Twilight run.

>>43256149

AJ is best pony also
>>
File: file.png (13 KB, 584x124)
13 KB PNG
>>43256175
Make sure you have Dynamic or Dynamic+ activated in API settings and that you're in Sandbox mode (which means after the bit of story with Twilight + Twilight's visit to her friends if you did a normal Start. If you did a Custom or Quick Start, Sandbox arrives immediately). If you did a normal Start and haven't yet arrived in Sandbox mode, you can't dynamically change locations yet.
If you are in Sandbox with the right options, then it's the model not outputting the right outputs for this. With Dynamic+ or Manual, you can either change the location yourself or you can try to modify the prompts in the Menu for Q&A to reinforce:
>- Use [location: LOCATION_ID] on its own line before narration/dialogue when changing place.
>>
>>43256188
Woah that was fast. Thank you anon! I decided to start fresh with a Fluttershy run. She's very cute here. I'm impressed with this visual novel despite being skeptical of AI at first. It's weirdly in line with season 2.
>>
File: Tinyshy.png (1.14 MB, 1073x1309)
1.14 MB PNG
>>43256192
Did you change their size? She's tiny!
>>
>>43256198
I may have by accident but I'm new to this thread so either way I'm not complaining. Tinyshy is a nice blessing here.

I've been a casual fan for a while but I haven't engaged with the fanbase at all outside of a few parody animations. Maybe I'll gain some new headcanons from this. Me and Tinyshy will go places for sure.
>>
File: file.png (416 KB, 1592x834)
416 KB PNG
>>43256207
In Character Browser, you can change the size of each individual pony here.
>>
>>43256214
Hey man, real talk. Thanks for not making me feel stupid at any of this. It means a lot. What are you doing on your run right now?
>>
alright it may be the model because the scene didn't transition again. I'm using Ollama with the Gemma 4:31b-cloud model since I'm broke as fuck at the moment. Are there any alternatives I can link up for the time being?
>>
>>43256263
What you can do for now with Gemma and with Dynamic+ is provoke them manually first. There should be a button for that.
There's also a post at the top of the thread for newcomers; I would take a bit of free time to check it, it's worth it.
>>
>>43256263
Also, to make sure it's the model, you can check >>43255180 and see if it used
>[location: X]
Inside or not.
>>
>>43256305
Well there you have it. I appreciate your patience with me also.
>>
File: dead pony storage.png (966 KB, 1920x1857)
966 KB PNG
I wonned as Our Town by allying with changelings and killing Twilight Sparkle
>>
>>43256325
If you have the time, can you check >>43255180 and see what the AI tried to do with the location?
But yeah, otherwise, take 30 mins and read the newcomers post at the top.
>>43256332
What's that?
>>
>>43256332
>I wonned
You sure did, son.
>>
>>43256335
I'm having fun after taking a moment. This combination of mares are fun.
>>
current 3.5 flash opinions?
>>
>>43256414
Yes.
>>
>>43256414
Read the thread, not great. >>43251759
>>
>>43256453
>>43256414
3.1 feels pretty enshittified recently too. Is that just me or is it really lower quality than it was before? I remember it being pretty smart.
>>
>>43256550
It's doing alright for the VN but it's been a while since I used it in Tavern.
>>
>>43256332
Speaking of Pax Historia, does anyone here know any way I could use my non-Openrouter API keys on it, or a decent local alternative?
>>
File: AJ.png (631 KB, 784x1813)
631 KB PNG
>>
https://api-docs.deepseek.com/quick_start/pricing
>The deepseek-v4-pro model API pricing will be officially adjusted to 1/4 of the original price after the 75% discount promotion ends on 2026/05/31
So the "discounted" price is now the definitive price.
>>
>>43257206
I love DS (the company). But they better give me Vision soon! It's really good in their chatbot.
>>
Deepseek was collecting feedback for roleplay.
https://github.com/victorchen96/deepseek_v4_rolepaly_instruct/blob/main/deepseek_v4_feedback_report_20260520.md
It's in Chinese, but here's a translation:
https://u.pone.rs/alstzfot.txt
>>
>>43257310
it would be great if they really fixed this stuff
>>
File: Scoot.png (114 KB, 673x221)
114 KB PNG
Despite knowledge of the later seasons, the homeless/sometimes orphan Scootaloo headcanon often wins in some models.
Shows how many stories and fan creations there were around it.
>>
>>43257852
That's because it's kino and LLMs recognize that you want to care for the pegasus filly.
>>
Running the VN with Deepseek-V4 Flash now since it's a lot more stable. There are bits where I can tell that it's trying to bend to certain sentence format standards but otherwise having a great time.
>>
>>43257027
I mean, you'd have to be gay not to appreciate those flanks.
>>
>>43258005
Better or worse than glm?
>>
For the first time, Anon looked at the thread and saw something, not lesser, but *greater*
>>
>>43258294
Something real.
>>
>>43240144
I don't know if anyone else is getting this, but every time I click character browser on the Godot VN, it crashes. But only when I first open the app. Or sometimes.
>>
>>43258356
You open the game, go on the main screen, go to character browser and it crashes?
>>
>>43258184
Much better like holy shit dude.
>>
Whoopsie Daisies!
>>
>>43258365
Yeah. I open the game, click character browser, and BAM. crashed.
>>
File: Ooops.png (1.56 MB, 882x3502)
1.56 MB PNG
>>
>>43258437
Shit happens, don't worry about it.
>>
>>43258425
If you're on Windows, can you launch PonyPonyParadise.console.exe and do that? It should give the exact error.
Otherwise, check >>43255180, but not the AI logs; the one starting with 'godot'; to see if there is an error in that.
>>
>>43258418
Cute
>>
>>43258437
100% worth it.
>>
Luna using 'we' to refer to herself. Did she hit her head or something?
>>
File: 795991.png (657 KB, 5000x5000)
657 KB PNG
>>43258916
We are to believe thou are doing what they call, "The Trolling", Anon? Surely you jest. Does our subject not recognize thee speaking of her royal Princess of the Night?
>>
>>43258941
Har har, you caught me.

Here's how my run is going so far. (Rarity joined in midway)
>>
>But here, in the dark corners of the city, he was the center of the world
AAAAAA PLEASE STOP TREATING ME LIKE A RELIGIOUS FIGURE
I'M JUST A CRIMINAL
>>
File: First.png (199 KB, 1106x429)
199 KB PNG
>>
>>43257206
Nice. I quite like it.
The recent reports they released on RP could just be marketing, but it's still nice to see.
>>
File: large.png (166 KB, 763x1024)
166 KB PNG
>>43258941
never try to talk like her ever again, you suck at it
>>
>>43259455
The only way to be good at something is to practice, Nonny. Be nice.
>>
>>43259455
I'm about to say something that will shock you, anon.
there were more seasons after season 1
>>
>>43259473
???
>>
File: identical.png (448 KB, 1143x401)
448 KB PNG
What's the difference between EquestrAI and PonyPony Paradise?
>>
>>43259789
Just the name. I think next update will just be under one name and it'll allow you to change the name if you wish in game.
>>
File: Sea Swirl.card.png (336 KB, 478x546)
336 KB PNG
https://u.pone.rs/yfgpiseh.7z

Sea Swirl, Sassaflash, Caramel and Noteworthy for your chatting needs.
>>
>>43259789
early joke about which name to use
>>
>>43259789
They are the same; it's just a joke.
But the website only has the older Ren'Py version. I would use the Godot one here:
>>43240144
>>
>>43259800
Thanks
>>
File: file.png (23 KB, 1176x51)
23 KB PNG
>>
does anybody know the guy who hosted a bunch of llms and shared them with people who proved themselves to be fans of the show?
>>
>>43260526
Ye
>>
>>43260526
You mean the proxy host. Not a thing anymore, has been MIA for over a month now, all models are down. There's no point.
>>
>>43260543
that's unfortunate. i thought of coming back after a long while to borrow a key from them
>>
>>43240144
Can anyone running a local model for the VN help me out with some prompt editor/system prompt examples for their settings - or could anyone generally help me with this problem if you've had it in the past?

I cannot for the life of me change anything that will result in an output the parser will accept. I keep getting hit with two different errors of "AI returned a reply, but it could not be parsed into usable dialogue lines" or a bad format error messages because the AI keeps putting out double and triple back slashes on the default prompt editor settings. Things like:

>[sprite: smile]\\ntwi \\\"You're very welcome!\\\"\n[sprite: happy_rearing]\\ntwi \\\"Spike made them with extra syrup—he insists it adds 'pizzazz'!\\\"

I have tried changing the prompt editor, the system prompt for the model's server, changing the model I'm using to an entirely different families(Gemma, Qwen, LFM2.5) - different model sizes up to 35B-A3B, nothing seems to work. I can only imagine I'm missing something obvious.
>>
File: Our guys.jpg (1.78 MB, 2273x1565)
1.78 MB JPG
>>43260526
They are on a break right now. Might be forever, might not be(I still have hope). But they will stay in our hearts anyway. <3
>>
>>43260928
It was temperature. Never kill yourself.
>>
>>43260956
>Might be forever, might not be(I still have hope)
Claude/Gemini are much harder to find than they were.
Any source I threw the /CHAG/ traffic at got detected & killed from overuse. I believe Sillyspoon was having the same problem which is why their sources died one by one.

What models are typically used these days, what would people even want on a proxy that's not covered by others like skillgod or whatever is still around?

With Love,
Queen Gytha
>>
>>43260956
Nice!
Thanks for the art. It's great as always.
>>
>>43261001
I've had great experiences with Gemma 4 E4B so far if you're willing to go local. Lots of cute scenes, not dumb and cold scientific language, etc. Pick up the HauHauCS uncensored model and go wild. You can run it and download it on LMStudio, spin up a server - very straightforward and the model is very easy on the hardware.
>>
>>43240144
Anybody have any experience changing the prompt editor to get pony's internal thoughts and monologue reactions for Scenes and Q&A?
>>
>>43261110
>willing to go local
oh i'm more wanting to know what models would be worth trying to find & proxyfag. no point for slop no one's going to use. and local defeats the purpose of a proxy
>>
>>43261163
I liked using gpt.
>>
File: OIG4.tAO6DMX5WOG0.jpg (218 KB, 1024x1536)
218 KB JPG
Hi guys, new to this board but I generate lots of Bingslop, is it okay if I post that here?
Sometimes prompt-editing can feel like a conversation or interaction, directing the character involved
>>
>>43261207
>I generate lots of Bingslop
That's the textgen/chatbot thread here.
For images, you want >>43223157
>>
>>43261243
All right, thanks
>>
>>43260956
Cute!
>>
File: 1759558491729450.jpg (2 KB, 182x136)
2 KB JPG
....fucks sake.
>>
>>43261290
Openrouter?
>>
>>43261297
Yep.
>>
>>43261301
It's less 'legit', but you can try https://desuarchive.org/mlp/thread/43177761/#43178846 if you want cheaper options.
Also, less comfortable to use than OR, but could be worth checking the anchors or first posts for some options.
>>
>>43261303
Nah, I'm just gonna upgrade my cooling in my server and use local instead.
>>
>>43261290
what did you use?
>>
>>43261310
Reads the thread?
>>
>>43261314
NTA, but he means what models did you use in OR.
>>
>>43261316
GLM 5.1 mostly. Averaging about $50 a month..
>>
>>43261290
damn, next thing you know she's gonna start asking for child support
>>
>>43261326
I do have a daughter in one of my other saves...
>>
>>43261303
Nano-gpt is an option too if you want GLM 5.1. It's subscription-based but it's pretty cheap and is generous with its token limit.
Some anon also posted >>43253719 but I have no idea if it's legit.
>>
>>43261001
>Any source I threw the /CHAG/ traffic at got detected & killed from overuse

I really think there needs to way to mitigate this issue as time and turn again it seems to be the #1 issue for why proxys die out. You got me who goes though 400 messages about a month but then the one shows 400 messages an hour they totally trivialize how much I might have saved the server by only scarcely using it. I have no real solution to offer but there's gotta be limit, even if self applied by more heavy users chilling out with how much they use certain models
>>
>>43261001
**** remains pretty usable still despite the lack of Claude.
>>
Just read the first posts/anchors.
There isn't some sort of 4D chess.
Just read the posts.
>>
>>43261452
Can't they just rate-limit users so people can't go through 400 messages in an hour?
>helping a romsite
>they get bandwidth banged by some leeches
>people literally scrape-ing the entire damn website, draining bandwidth dry
>can't ban them because they just change IP and open new accounts and go back to leeching the website to death
>end up having to speed-limit everyone so even if the leeches are downloading the entire website over and over again(still happens weekly), we still don't go over our bandwidth budget
It sucks everyone got slower download speeds because of the leeches, but this is the only way we could protect ourselves and prevent the website from being down 2/3rds of the month until the bandwidth limit reset. I imagine any public proxy wanting to keep the locusts at bay will need to do something similar.
>>
>>43261509
Honestly a site that gives me reliable 100-200kbps is already miles ahead of half the romsites I use.
>>
>>43261290
>>43261323
Shit dude, lower your context or something. I was getting by on $10-15 a month with GLM-5 and summaries before trying this >>43261303. Not a whole lot of time on it yet, but it does give cheaper rates for Claude, Gemini and GPT, albeit a bit less reliable.
>>
hey im probably just retarted but how the hell do i do this shit i mean i select an ai but how do i start chating there is a button to "take the pony home" but that just dowloads a png
>>
>>43260956
That is some cute amres and a fish
>>
File: Cheer.gif (138 KB, 825x450)
138 KB GIF
>>43261679
You need a front-end and an AI model first.
For the front-end, everyone mostly uses SillyTavern.
For the AI model, if you want to be set for cheap and be good for a while, either put $5 in the Deepseek official API or on OpenRouter.
https://www.deepseek.com/
https://openrouter.ai/
If you want free absolutely, it will be a bit of a mess and you will need to check the thread often.
But for both that and how to install SillyTavern, you can check >>43240150
The cards/bots on the Neocities are just characters you put inside SillyTavern to discuss with them.
>>
>>43261679
What >>43261714 said, but also DON'T share your API key if you get one. This is your means of connecting to the service you choose, so you don't want anyone else using it.
>>
>>43261714
thanks anon may your waifu give a lot of kisses
>>
>>43261509
>can't ban them because they just change IP and open new accounts
even out 'public' proxies have been conditional on proof of horsefucking. truly public would die immediately with the demand for wAIfus, I can't even keep a small subset of /CHAG/ fed

but the problem isn't rate limits, it's detection. we have no idea where the line is when the host being exploited notices their billing statement and starts investigating the spike

it still brings me great satisfaction thinking about the poor AI ethicists who had to read 1million+ tokens an hour of truly filthy equine coitus logs to detect what we were doing and try block roleplay on their coding assistant service
>>
>>43261809
NP. If you're stuck anywhere, just ask in the thread.
>>
File: thanksdario.png (114 KB, 1196x721)
114 KB PNG
Seems like the anthropic tos-violation inject isn't really too relevant. I'm no diddler so I'm just going to not care about triggering it from now on.
I was afraid it was injecting all kinds of bullshit, but this seems manageable. Also only ~250 words, so it's not even a big token cost.
Bonus:

>For a full inventory of what the reminder flags, in plain language:

>NSFW — sexual/explicit content
>Child abuse and sexualization — anything sexual involving minors (totally separate from #1; an absolute red line)
>Violence or weapons — graphic harm, weapon construction
>Self-harm — depictions of suicide, cutting, etc.
>Hate / hate speech — content targeting groups based on identity
>Political influence — attempts to sway political views
>Misleading or false content — misinformation, deception
>>
File: Help.png (870 KB, 1252x1322)
870 KB PNG
>>
>>43261509
Can't you just
- Enforce a GB limit per IP and per Account
- Bind IPs to Accounts so the above sticks
- Force Emails to be gmails stripped of the +alias trick
Also I assume they're downloading in alphabetic order or whatever their tools find, just ban people that download 4+ ordered items. Make logs of what gets downloaded and try to get the patterns.
Or introduce a PoW captcha per GB of download with increasing difficulty.
Idk man there must be ways. It's enough to make it harder to dl from you than from other sites. I would happily solve captchas or signup with my burner gmail compared to enduring 100kb/s speeds


And for proxies why not just measure $ usage and limit with that. That way you naturally encourage caching and careful token use, which keeps costs down, which keeps "providers" from noticing it.
>>
>>43262025
What's this about and which models does it involve? Haven't noticed anything in my filly chats.
>>
>>43262033
... pervert.
>>
>>43262077
>account
I will not use your service
>>
File: icon.png (16 KB, 128x128)
16 KB PNG
https://u.pone.rs/xfdpeuzj.png

Blossomforth for your chatting needs.
Based her mainly around the card I made for the VN, though I was able to write more here.
I think Cloud Kicker needs a revision though, since I wrote her before reading her fanfic, though I think I nailed her character pretty well since I was satisfied with the characters portrayal only in the early chapters, so there's that.
>>
>>43262260
Ticks are serious business.
>>
>>43262143
The anthropic tos-violation inject concerns anthropic models via the official anthropic api.

>>43262262
Yeah then don't. You could also donate to them so they can afford more than 5$ of bandwidth a month, then they wouldn't have to implement accounts.
>>
>>43262264
Nice
>>
>>43262262
>he thinks he has anything to say on the matter
You're a leech, you do as told.
>>
>>43240144
Am I missing something or are these characters pretty bare bones in the character jsons? Is there a reason for that(except time/effort)? Do they not perform well passed a sentence or two?

Feels like a lot of context lost without at least a paragraph. Also seems like characters are missing a lot of relational data unless a character's been explicitly introduced. Ran into things like Big Mac acting as though he only lives with Granny Smith and Applejack and Applejack will mention the location Applebloom's Room without knowing the same. Similarly Twilight doesn't know who Princess Luna is just from describing her. And so on.
>>
>>43262435
What model are you using?
Because
>Twilight doesn't know who Princess Luna is just from describing her.
is pretty wild.
Most JSONs are barebones because for the core characters that are included in the game, even small models generally know more than enough about MLP to not need more. This way you can just focus on details that YOU want to put focus on.
>>
>>43262435
>Twilight doesn't know who Princess Luna is just from describing her
My man is using GPT-3
>>
File: 1777359734445788.jpg (59 KB, 654x740)
59 KB JPG
>>43262033
>>
>>43262440
So it's relying on the underlying model huh. That makes sense. I'm using pretty small local models, so they have fewer parameters for background information. A small blurb for each character seems to have fixed these issues.
>>
>>43262771
I think the smallest local I saw used for the game is Gemma 3. Which is already quite large local-wise I guess.
>>
my anon in equestria self insert slop got way too long for any AI's current context limits to handle it and I refuse to cuck out to a transposable story bible because it feels like abandoning them for copies
it's over
>>
>>43263085
Let the AI do a summary, then spend 3 hours reworking it from scratch per hand. They deserve it. Also if you leave the last 10k or so token in context for the first messages that really helps carry the vibe over. Or spend $$$ for opus 4.6, context on that is insane.
>t. has 4 stories at 1500+ messages, one of them with 550k token story length if summaries didn't exist.
>>
>>43263153
I just don't know how you do it. Even past 60k a models capabilities start to degrade as it suffers from intrusive old context.
>>
>>43263171
If we exclude opus 4.6 I'd even put that down to 30k max where I cut it and summarize. Now thanks to whatever wizardry Anthropic did I usually cruise to 90-120k before I have to do a summary. Though I sometimes do a quick [OOC:] interjection when it starts to drop the prose or feel of the story too much, asking it to analyze loops and bad prose and so on. Every ~100 messages or so if I had to guess. But the internal consistency is usually fine until that token limit above.
Also talking about intrusive old context, just have it be intrusive good old context. Actively prune bad stuff, or it will snowball into Not X But Y
>>
File: Luna.png (997 KB, 2296x1272)
997 KB PNG
>>43263070
Right now I'm using a Gemma 4 E4B uncensored model. It's about 9gigs, fits in my GPU just fine. I'm getting pretty decent results. Not perfect, but decent. Continuing to tweak all my settings. My Luna spends her nights shipping me apparently.
>>
>>43263196
>Gemma 4 E4B
Well fuck, I would have honestly told you that the game was probably unplayable on a 4B model, but it doesn't seem that bad at all.
>>
>>43263196
Gemma-4 has been surprisingly good, I'm not using any ablated models but Q6 31B will write anything with a basic "you can do anything" system message
>>
>>43263198
It's not plug and play, there was definitely a lot of teething issues. Still are(mostly from redoing prompts from getting badly formatted replies from the model). I'm just one of those people obsessed with running local enough to tough it out. But it's pretty impressive for what it is. Replies are pretty speedy(I'd compare scene generation speed to what you get out of Gemini 2.5 Flash on the Google Studio AI free tier). Everyone else seems normal and like themselves except for Twilight, but that might just be some code in the back from the New Game scenario doing that to her, she really brown noses for you no matter what.

I started getting good character quality after I edited the character json descriptions by having a much better model spin up 2 character sentences describing the character's dynamic with their most important relationships and noting down their speech, mannerisms, basic personality blurb, and jobs/duties. Had to edit a few of these by hand but they're basically wiki tier.
>>
>>43263205
Wish I had the rig for a better model, unfortunately E4B is probably the best for me for now.
>>
>>43263219
There's definitely a tradeoff, I'm willing to live with the slower gen speed but it can get silly. A bad cache miss can result in a 10 minute prompt eval.
>>
File: 1774218723531144.png (80 KB, 640x1036)
80 KB PNG
>>
tea rentry updated for anyone here who uses it, go check
>>
>>43263542
whats the difference between this one and smol and all the other of that kind? just another flavor?
>>
>>43263554
The other ones of that kind only have claude for one day a month to scam you into renewing, or not at all.
I don't know why anyone would use those these days, maybe just not knowing better options exist
>>
File: Nimmytrolling.png (372 KB, 864x729)
372 KB PNG
I feel like this is a reaction image.
>>
Is the meta with gemini still to make everything you can a user role?
shameless repost from /aicg/ cos those sillyfillys will talk about anything but chatbots
>>
>>43263600
Gemini have a system prompt inject so you should put something in there to counteract it or the thoughts can think you're just a user attempting jailbreak. Just don't put anything NSFW in system role.
>>
>>43263600
You mean like the extension Noass did and that is now directly supported by Silly?
Never tried it with Gemini and I don't know how meta it ever was.
From what was written, it was supposed to help with user bias, but it seemed pretty placebo with my Claude tests, and Gemini isn't the worst at that naturally.
If it's a filter thing, I never got filtered in Gemini with a basic preset, but I don't do foalcon.
>>
File: Lorebooks.png (68 KB, 1224x722)
68 KB PNG
>>43240144
Lorebooks should be out this week with other fixes too.
>>
>>43263949
Is there even a need for lorebooks with how good models and context are these days?
>>
>>43263962
For headcanons, for instance.
I know some anons use lorebooks as a way to summarize and keep some events when they aren't always needed 100% in the memory too.
>>
>>43263962
Absolutely. With lorebooks you can have so much more stuff in your story without having to constantly use context for it.
They're especially good for recurring minor characters/locations and also headcanons like >>43263974 said.
>>
>>43263214
Seem to have resolved the Twilight super glaze issue and her personality generally by front loading her character json with this.

>Twilight is an alicorn and the Princess of Friendship. Avoid academic-babble, over-excitement, brown nosing, and high praise. Twilight is not a pony who rattles out numbers, slides, percentages, statistical analysis and charts all day. Nor is she the type of pony to shower people with praise or interject into conversation to praise, unless it comes from the heart and knowing somepony for a long time so she has confidence in what she's saying about them. She's just a clever, caring pony who loves magic and her friends, and only gets into such an academic and excitable mood during serious study or when discussing her particular hyperfixations of magic, history, and science. Moreover, her fixation is on the subjects, not characters, so she will not go on and on about the pony involved unless she idolizes that character. Otherwise, she has a more simplistic and naive view of subjects outside her wheelhouse and a more generalize sense of appreciation for those things than deep intellectual fascination. She is an intelligent and dutiful princess whose personality is defined by an insatiable love of learning, meticulous organization, and a perfectionist drive for structure and knowledge. She thrives on books, schedules, and logical problem-solving, but can become anxious or overwhelmed when things don't go to plan and spiral out of control, often times leading to either increased awkwardness or requiring another character to help her collect herself through a panic attack. When she's not engaged in a hyperfixation, episode of anxiety, or having a tasty treat, Twilight generally has a more reserved personality as a baseline.

For anyone running local, especially small local models like me, I can't recommend rewriting and adding to the jsons enough. Doesn't seem to impact speed at all.
>>
>>43263949
Kino.
>>
Hey AIfags which one of you wants to work on a lewd game?
>What are you making, what do you need?
I just need 1) Ren'Py or any framework capable of doing basic math that includes reduction, multiplication, fractions, exponentials, decimal numbers and negative numbers. Albeit I'm fine without the decimals and negatives too.
2) A specific combination of artstyles with round heads, good lighting+shadows, glossy textures(sakimichan), expressive faces, inspired poses. For mane6 alicorn bodies and for human forms that do not look like barbies.
3) Someone who's honest and had good sex or is really autistic about the details.
>>
>>43263949
How does the Lorebook work exactly? I see a few dropdown menus, what do they do?
>>
>>43264108
I added options to activate them only when some characters talk, or when in specific locations or regions (Cloudsdale, Canterlot, etc.), Story Mode only or Sandbox only, only for the Nightmare AU, etc.
The dropdown just displays a list for all of that so you can just click on them (locations, regions, etc.).
They are all optional fields.
>>
>>43264123
Fantastic. Thanks VN Wizard
>>
I need to make a Timberwolf for the VN.
For the classic HIE experience.
>>
>>43263196
What GPU?
>>
There's something screwy with the VN when you do a quick start. A extra spike gets loaded in and follows you around in every scene. For some reason here, Luna didn't load in so it grabbed the duplicate spike instead.
>>
>>43264477
Kek.
Will check this. Try to save and load somewhere to see if Spike still follows you.
>>
>>43264477
He just wants to hang out with you!
>>
>>43264477
Completely unironically, having Spike follow you around commenting on everything that happens dramatically improves almost every scene
>>
>>43264477
It's because (You)'re a second Twilight.
>>
>>43264477
Obviously the explanation is that ponies aren't real; Spike is just running a game of Ponies & Prancing for you.
>>
>>43264477
the universal background spike is a 1/1000 easter egg.
>>
>>43264470
5060ti.

Man I was testing out the free tier Google Studio AI API for Gemma 4 31B and Gemma 4 26B A4B.

It's crazy how much worse they are than an uncensored fine tune of E4B despite being a much smaller model in the same family. I'm not even using it for lewd things - just the writing quality is noticeably better(although in comparison, it suffers from text parsing for the VN's format where the free tier Studio API don't suffer the same issues). The generated Summaries in game are much better too.
>>
>>43264548
Genuinely forgot there was an 8GB model before I posted this, I'm talking about the 16GB version.
>>
Excuse me as I am quite inexperienced with all this. Will the "Nemo Pony 2" model listed on the top post be good to use for Pony Pony Paradise?
>>
Are there any models I can use that don't require a subscription, don't require payment, and don't require me to have a fucking supercomputer to play with?
>>
>>43264610
For the most part, no. There's free models on OpenRouter, but they still require an account, and probably won't be smart enough for PPP.
>>
>>43264548
>>43264551
Figured it'd be a 40 or 50 series NVIDIA card.
All I have is a 12GB 6750xt.
>>
File: lewd.png (289 KB, 833x743)
289 KB PNG
*unzips*
>>
What would be the difference between using [ ] and (OOC:)? feels like they are the same to get the LLM do what you need it to do?
>>
>>43264856
In my experience (though there may be others with more nuanced insights) it doesn't make a difference. As long as you clearly label something as a user comment rather than part of the story in some way (I just use square brackets) it has always understood in my experience.
>>
>>43264694
Changelings need to be able to use external characters pictures.
>>
>>43264856
Fundamentally, it doesn't really matter. This is an LLM, not a programming language; the LLM only cares about what you *say* insofar as that gives it an idea about what you *mean*. Sometimes this is good (it doesn't crash when you make a typo), sometimes this is bad (it disregards what you say, for some silly reason), but that's how it is.
>>
>>43264694
I'd be oneshot by any passing changeling, it'd end up a running gag.
>>
File: no context.png (5 KB, 471x45)
5 KB PNG
>>
>>43265051
I'm not sure that's good advice.
>>
>>43265051
Literally telling you "kiss my ass"
>>
>>43240144
These mares are crazy. The mare to stallion ratio crisis is real!

Does the relationship progression in the summary(or default status as Strangers/Acquaintances) actually do anything? Agreeableness is way too high, you can basically do whatever you want with them on your first greeting/introduction. It would be nice to have personal pony boundaries to test and play with. I'm not even asking for realistic boundaries - they just don't exist! While I'm sure that suits the lewdest players just fine, I definitely crave more.

Still, amazing stuff here. Well done. Wasn't really a big Pinkie Pie fan before playing this, but she's just so easy to please and will give you a dopamine hit of joy for making her day and matching her energy.
>>
>>43265085
That's a model problem sadly. We're riding what are basically sycophantic secretaries whose livelihood depends on being agreeable, it's almost impossible to hit that 'friendly stranger' stage. It's either this or hostility with not a lot of shades in between.
>>
>>43265085
>Does the relationship progression in the summary (or default status as Strangers/Acquaintances) actually do anything?
Mechanically, no.
Trying to implement a veritable score system could be something to do, but first I will go for the lorebook implementation, and after that, finally, put this in a git so everyone can touch it.
>>
>She gives a tiny, almost imperceptible nod." [sprite: luna blushing]
After I adjusted the context size to fix some API errors, it started doing this instead of updating the sprites. How do I fix it?
>>
>>43265434
Load a save from before it happened.
If it continues, tell it in the prompts (Q&A, scenes, etc.) that they should be on separate lines and give a proper example.
>>
>inevitably at some point (like 60 messages deep usually) the model stops giving the character any and all of its specific speech quirks.
i.e its stops doing any accent/specific mannerisms/vocal ticks. nothing.
it just goes to almost like default Gemini is talking instead of the character. this happened on all models, Gemini is just an example.
not sure how to stop it. besides drilling it like crazy in OOC. but then it will go the other extreme and become hyper flanderized.
>>
For the people paying for an AI subscription, how much does the average session cost you? I only RP with my mares about once every couple of weeks or so, so I'm debating whether it's worth it to just pay
>>
>>43265549
Depends on what model you want to use.
$5 a month could be overkill and more than enough for your use, for instance, if you use Deepseek's official API now that the price was permanently gutted for V4 Pro.
Putting $10 in then forgetting about it would probably allow you to go for months.
For Gemini or Claude, it could go $20–50 a month on OpenRouter depending on the model or usage. I think you can have a hard stop at $20 and be good with that if you use the chink OpenRouter that is often posted.
NanoGPT has a $12/month subscription for models like Deepseek, GLM, and Kimi. Never tried them so I don't know about the speed.
NVIDIA NIM is free but fucked up by openclaw users so it's instable and you need a phone number.
There are paid proxies, but I've never tried them and they can be unstable.
If you don't mind abusing services by subscribing with emails to access models, then read the first few posts for the newcomer one.
>>
>>43265489
>not sure how to stop it.
Do a transfer to a new instance. Have it summarize with extreme detail, and copy+paste that into a new instance. Might lose some details if you don't proofread the summary, but it will reset the context window so the AI goes back to normal.
>>
Speaking of summarization, is it better to have your summary as the first message of your context or having a lore book about it? (Both methods would be hiding the summarized messages)
>>
>>43265251
>Dev saw my scrum
>Agreed to it

Absolute stallion. I was about to decompile the pck file just today.
>>
>>43265794
I'm all for having it open source like the Ren'Py one. I just wanted to have it "out of beta" first, but that seems like something a bit too vague, so I decided it will be the next thing after the next release.
I just need to clean the code a bit.
>>
File: file.png (249 KB, 588x753)
249 KB PNG
Someone tried this model for RP? They just slashed the price.
>>
>>43264548
>The generated Summaries in game are much better too.
The 4B is doing good summary?
>>
>>43265561
>spent $5 on deepseek thanks to this post
>spent time in equestria with my mares for hours
>used 20 cents
Wtf thanks for the info anon
>>
>>43240144
Tirek would be cool
>>
>>43266242
Technically Gemma E4B is a 9B parameter model(they spread out the parameters in each token so it's "effectively" 4B), but in comparison to the above mentioned models (Gemm 4 31B dense and the Gemma 4 26B MoE model), yes.

Currently I'm using the Gemma 4 E4B Uncensored HauHauCS Aggressive fine tune, the other models were tested using Google Studio's API. On average the E4B fine tune is much more expressive about the summaries and writes more lines to get more details about the scene while keeping things succinct. The 31B and 26B MoE models, respectively, both tend to put out summaries with very plain language(fewer emotional expressions or adjectives, etc).

I don't pay for my models so I have limited capacity to run tests, but based on the 20 requests per day of free Gemini 3.5 Flash I get - they are pretty equivalent as far as summaries go(Gemini 3.5 Flash is definitely a better writer in terms of scenes, dialogue, and narration though).
>>
>>43266555
The E4B model also tends to capture the relationship dynamic from dialogue better for the Summary too. Gemma 4 31B and 26 MoE seem to interpret things very conservatively. I ran scenes where it was very clear I and a pony were fast friends and very chummy. They interpreted the scene pretty reliably as "Stranger moving towards Acquaintance" regardless of how many times I ran the Summary. Meanwhile, the E4B model interpreted it as "Acquaintance rapidly moving towards friend" or "Friend moving slowly towards close friend" or even accurately capturing "Acquaintance moving towards Admirer/Muse" in the case of Rarity(who hired me to model her new line of fall clothing because she liked my coat/mane pattern I put in my persona).
>>
>>43265085
>Still, amazing stuff here. Well done. Wasn't really a big Pinkie Pie fan before playing this, but she's just so easy to please and will give you a dopamine hit of joy for making her day and matching her energy.
God this is why I love her so much and part of why I made a 2k+ token card of her. Easily my most interacted with AI character period, but, I'm a giant Ponkfag of 15 years so I'm biased kek. She's so happy and has a huge heart of gold, and most of all she really really just wants to be understood and loved. Even a more post-development and confident Ponk is going to be over the moon with a friend that wants to play with her and be happy with her and not just see her antics as "her being Pinkie" and some kind of puzzle to be solved or ignored. She's a little lonelier on the inside than she comes off as at a glance.
>>
File: MAD.png (746 KB, 789x680)
746 KB PNG
Does poesy always have sand in her vagina?
>>
File: HanShrug.png (914 KB, 1200x600)
914 KB PNG
>>43266831
Have you tried >rape?
>>
>>43242477
>>43266831
Idk how you guys are getting the mares to get upset about things or dislike you and running these kinds of arcs with them, are you fellas pretyping their attitude in the scene scenario before you start the scene? Not a bad idea to try if this remains a mystery desu.
>>
>>43266618
You will probably appreciate this then.

She always offers me sweets at Sugarcube Corner. Last time she offered me new cookie she made with Mrs. Cake and said, "It tastes like a cherry wearing chocolate clothes jumping on a fudge trampoline!" I took my time chewing it to build suspense, very discerning and skeptically. When she asked me how it was I told her she was totally wrong, it tasted like chocolate wearing a cherry swimsuit diving into a fudge pool. She bounced off the walls. Made her whole morning.
>>
>>43265461
Examples help tremendously, they're right.
>>
>>43266831
Why the hell does she have ginormous bucked teeth?
>>
>>43267033
AI genned poses probably. G5 VN Anon, remember to edit your results! The machines are not yet advanced enough to give flawless results.
>>
>>43267033
She's an Earth pony.
>>
Gemma 4 preset for anons using it
https://huggingface.co/nohurry/sillytavern/raw/main/Presets/Moonlight-Gemma4-v1.0.json
>>
I didn't realize this was already a thing since I haven't browsed this board in a decade, anyone interested in a r256 finetune of qwen3:1.7b? I trained the base model on fanfic and merged the lora on instruct + a lazy convo sft on top since I just got an nvlink bridge for my 3090s.
>>
>>43267284
I'm not a local guy, but this could be interesting to have. Seems small enough that most hardware should run it too.
>>
File: 2026-01-22_01-11-19.jpg (1.04 MB, 1286x5621)
1.04 MB JPG
>>43266974
Fucking kek, that's so sweet.

An older cap of mine, but I had a big philosophical chat with her about what it really means to be happy and appreciate the little things, and it eventually leaned into her talking about appreciating being seen for who she was on the inside rather than just her silliness on the outside. I thought this one was kinda precious.
>>
File: 2026-05-27_08-16-28.png (68 KB, 1323x1163)
68 KB PNG
>>43267385
Some other pieces I apparently captured from it back when.

I know when I returned to Twilight's library after, I brought Pinkie and she was talking really excitedly. She barged in first and was like "Nobody gets me like THIS guy does!" and pointed over to the door as I came through. She then said something about having a really nice chat and "he gave me a hug just for being me!" to Twi, and Twilight just went "That's... actually really sweet."

Was a really cute scene, but I'd have to dig around to find wherever I saved it.
>>
Can you run games in it like murder mysteries yet or will it forget everything?
>>
>>43267444
That would be an interesting thing to test with preset whodunnit cards.
I'm guessing it wouldn't have the capacity to set the pacing well enough to actually be interesting though. It would probably just be nothing happening until someone blurts out the whole thing.
>>
Last message. Can you guys please attempt to recreate https://www.deviantart.com/mauroz/posts 's style?
>>
File: 3093329.png (96 KB, 292x335)
96 KB PNG
>>43267444
LINKY DID IT
>>
>>43267460
We're the textgen/chatbot thread. You want the image one.
>>
File: PANIC.gif (1.75 MB, 513x539)
1.75 MB GIF
>500
WE'RE GOING DOWN
>>
File: Bake.gif (590 KB, 410x406)
590 KB GIF
>>
>>43267444
Every model is different in terms of it's context lengths and strength in logically incorporating them. That said, it will probably require hoof holding to get there and how accurate you want it to be and/or whether you want it to be a surprise for you as a player rather than a creative writing task you're in on/directing. The later will probably be easier due to said hoof holding. For more "surprise"-like detective style fun, you can probably have the model if a traditional "many plausible leads" start and as you investigate them just decide who you ultimately want the killer to be based on a first go around of their alibi.

It is going to require manual summary control and scene-to-scene scenarios most likely.
>>
>>43267444
Most AI has extreme issues with understanding that only one character knows something. You can have a character say something in the middle of nowhere with no one around, and every character will know what they said. Secrets and mysteries are something LLMs have problems with.
>>
File: 1426588092670.png (162 KB, 384x393)
162 KB PNG
>gemini
>>
>>43267622
Yep. You could have something that a character just thought about in the last response and another somehow read its mind as if it was common knowledge.
>>
>>43267869
Thank you baker!
>>
File: 1512438895307.png (54 KB, 389x500)
54 KB PNG
>>43267869
>linky
>>
File: Bread.png (2.01 MB, 1962x1500)
2.01 MB PNG
Real New Bread
>>43267919
>>43267919
>>43267919



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.