File: adt logo.png (1001 KB, 2115x744)

1001 KB PNG

>>108859807
OP updates:
>Things removed:
Removed neighborhoods
Removed InvokeAI (dead UI and dead website)
Removed SDNext (nobody uses it)
>Things added:
Added Anime Style Explorer

>>108859866
Removed WAN2GP (nobody uses it besides WAN schizo)

>>108859866
>Removed neighborhoods
Based. /hgg/ does not have them either, they always lead to flamewars

>>108859807
Good bake.
>>108859866
All very nice.

>>108859866
Why not MUGEN???? https://huggingface.co/CabalResearch/Mugen

OP should have https://github.com/L33chKing/ComfyUI_LatentResidueCleaner

File: pine gun.png (3.59 MB, 1856x1536)

3.59 MB PNG

>>108860608

oguri feet...

thanks for bakering

>>108860133
/hgg/ is itself a dumpster fire

>>108860820
why would she do that

>>108860820
i want pine to shoot me

Whoever has reForge, can you check to see if there's an "edit" button when preprocessing via Canny on Controlnet? I need to see if a guide I found is legit. The button is nonexistent on Forge Neo.

>>108859807
Logo aside, what witchcraft managed to gen this?

File: file.png (178 KB, 357x387)

178 KB PNG

>melonpan

>>108861482
You'd be surprised how simple the prompt was. The firedotinc lora is pretty nice, look it up on civit if you want.

2girls, hatsune miku, kasane teto, spring onion. miku is poking teto's face with a spring onion. <lora:firedotinc_anima_kohya_low-lr_few_blocks_v4:1> , @firedotinc, <lora:anima-turbo-lora-v0.1:1>

Steps: 12, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Shift: 2, Seed: 1636268431, Size: 1152x896, Model hash: 14fffe8ad5, Model: anima-preview3-base, Module 1: qwen_image_vae, Module 2: qwen_3_06b_base-anima, Denoising strength: 0.4, Clip skip: 2, RNG: CPU, Original Size: 1152x896, Hires Module 1: Use same choices, Hires upscale: 1.5, Lora hashes: "firedotinc_anima_kohya_low-lr_few_blocks_v4: de7a983ba802, anima-turbo-lora-v0.1: 68ed0aec6ff4", Hires Shift: 2, Hires upscaler: Lanczos, Hires steps: 6, Hires CFG Scale: 1, Hires schedule type: Simple

>>108861782
>Steps: 12
>CFG: 1
Now this is witchcraft. I was assuming advanced techniques I've yet to find guidance for as most of my 2girls gens end in failure. Or is Anima straight up better in that regard? There's even a mote of interaction.

>>108858526
did you try rdbt?
>>108861782
>>108861833
anima wonnered.

>>108861833
that's because of the turbo lora.
however 2girls are usually very reliable with anima, what are you doing that it's not?

File: Screenshot 2026-05-20 011119.png (220 KB, 1242x1196)

220 KB PNG

>>108861482
There is. It only appears after you've run the preprocessor.

File: duo lick.png (63 KB, 1344x768)

63 KB PNG

>>108861848
>>108861857
Still on Illustrious until it well and truly goes the way of Pony. Only a select few times I pulled off 2girls but I honestly forgot exactly what I did with Forge Couple to pull it off except even the prompts had to be arranged very specifically. I did find an illustrated guide that combines both it and Controlnet but it's unfeasible unless you can draw "guide images" from scratch. Spoiler may not be enough for NSFW so here's a catbox.

Forge Couple + Mistoline: https://files.catbox.moe/jw3eit.png
Forge Couple + Inpainting: https://files.catbox.moe/b7b833.png

>>108861879
Not on Forge Neo, which has no Controlnet 0. Counts from 1 instead, but I suppose this'll be the second iteration of Forge I'll have to install this week. Neo was a pain in the ass due to asking for a specific version of Python. Had to uninstall 314 in lieu of 313. If I have no interest whatsoever in Anima, is it safe to "downgrade"? And does minipaint work in Re?

>>108861919
Python installations for AI shit are generally done within a venv (virtual environment) specific to the software. And yes, minipait does work with ReForge mut ReForge has been abandoned for a while now. I'm mostly using Neo these days, only reason I still have Re is I've been too lazy to migrate shit.

>>108861919
ah well I remember needing regional prompting - which wasn't exactly trivial - to get 2girls working on SDXL.
the nice thing about anima is you can just describe it and it will mostly work.

>>108861954
>ReForge has been abandoned for a while now
And I've been using Forge long after it was abandoned. Didn't even know that until a few weeks ago and was told that pretty much every fork is considerably less bloated. The UI certainly looks a bit cleaner on Neo but I had to get another fork of Adetailer. Rest of the extensions from the old Forge fortunately still function, but I wish I know about minipaint sooner. Must've been a Godsend for those who use controlnet and inpainting frequently.

>>108861879
Forgot to ask this on the onset but does that button also work on your end? I understand from another tutorial that Photopea pops up and lets you change the lines (assuming the lines show up properly), but it kept freezing when doing it on the old Forge.

File: 1771387981054669.jpg (385 KB, 896x1152)

385 KB JPG

>>108856928
>>108856914
Konan made the Anima one.
https://civitai.red/models/2494484/akihiko-yoshida-final-fantasy-tactics-style-anima-lora

File: 05678-871792032.png (2.55 MB, 1728x1344)

2.55 MB PNG

>>108861848
rdbt (0.35 at least) is really good. Honestly my initial impression was that it can do a lot of details but is kinda locked into a semi-realistic style but you can use the turbo lora with it to get a good balance between details and style adherence. Also it is via some sorcery better at following natural language than base model. The last image from previous thread with Chen looking from the ceiling hole which I generated worked on almost every seed whereas on anima 1.0 it only worked sometimes, the vast majority of the time it placing Chen in the room instead of above it or just making a complete mess with no Chen/partial Chen.

The one thing I can't really seem to run away from if using turbo is a million fucking sweatdrops on skin in a lot of gens, especially anything that involves nudity. Can shoop them out easily with healing brush tool but kinda annoying.

>>108861954
Neo is cool for mostly everything but for some reason it's missing the DDPM sampler which I as the only person in the world seemingly use (I did a fuckton of testing and preferred it over all else). If I had to describe it it's sort of like euler but better details. So because of that I still have ReForge for Illustrious/Noob use.

>>108862120
what is rdbt?

File: requiem_for_a_VRAMlet.png (10 KB, 1024x1024)

10 KB PNG

>oekaki left the room
It was a pleasure meeting you, Oekaki bro...
You were there through the good times and bad times of /adt/.
You saw it born, you saw it fall, you saw it stuck in limbo.
You taught us that we don't need more than 1024px to make KINO.
If it's January and you're not here, I want you to know that I wish you a happy 35th birthday...
I hope to see you again...

>>108862331
>1536x1536
it's not the same...

Right now I currently have 32 custom nodes, is that a lot or a little? If I take screenshots of my custom nodes folder, can you help me debloat it?

>>108862191
https://civitai.red/models/2356447/rdbt-or-anima
It's the WAI equivalent but for Anima, CFG distilled, more stability in exchange for aesthetics. More mainstream AI anime look. It's a good option when you just want to generate cute anime girls and aren't looking for a specific style.

>>108862391
isn't there a wai anima? why is this the "wai" for anima instead? i don't get it

>>108862463
What I'm trying to say is that Anima to WAI Anima isn't as watered down or sloppy as Illustrious to WAI Illustrious is right now. It's more like how Anima to rdbt works, at least from my own experience,

File: 00269-3339566607.png (1.1 MB, 1600x1280)

1.1 MB PNG

>>108861782
always feels weird seeing people use loras trained on my stuff randomly exist in threads I'm in
(not that I mind at all, probably closer to the reverse, still feels weird)

File: silver-haired riding hood.jpg (1.83 MB, 2400x1792)

1.83 MB JPG

I don't think she wants to escape.

the slop convergence is upon us.
entropy soon. very awa true to oekaki-saar.

File: grid-0013.jpg (1.13 MB, 3328x2432)

1.13 MB JPG

Against my better philosophy, I opted to download GIMP (because there's no such thing as a soft brush in mspaint) and use the least worst guide on controlnet I could find and proceeded to wing it. Results proved surprising in both ways.

Good: It pulled off coke-bottle glasses far better than the inpainting attempts earlier in the day. Granted, I'm probably barely scratching the surface on what controlnet's capable of but I expected far more aggravating time.
Bad: Contrary to some examples I've seen, it can't lock down text (I added the "body writing" prompt), so I'm obviously missing something the guide assumes I have or know. Also, the shirt's supposed to stay black. No idea what made it white half the time.

The least worst guide in question: https://arcenciel.io/articles/27

>>108862794
>very awa
That a NoobAI thing? Every checkpoint I looked into never has that among the recommended quality prompts.

>>108862769
Can you do one of Teto angrily biting the onion because she's had enough poking?

>>108862814
yes. grok has spoken.

So I want to use a character that has like 100x more hentai coomer art than official art, and whenever I prompt it in Anima every gen comes out with some exaggerated giga booba body or super coomerized proportions
Are there negative tags I can use in to force a more normal body type? Like should I neg stuff like "hentai, gyaru, @trex", or whatever? If so what tags would you use to make it look more like normal body proportions?

>>108863049
Official art and official style isn't helping

>>108863056
did you add the characters series in the prompt? grok has spoken.

>>108863049
just say medium breasts or small breasts. also try seeing if the official depiction has an artist tat you could use

>>108863049
I use these:
>small breasts
>skinny (with ribs in negatives)
>slim legs
>(wide hips:-1) (with negpip)
>(huge ass:-1)
Depending on what body part comes out too large.
There are other negatives you could use like fat, plump, obese, thick thighs, but since these have actual opposite positive tags like skinny and slim legs, I prefer those instead.

>>108863144
>with ribs in negatives
HUGE faggot

>>108863071
Yes.
>>108863108
I wish it were just the boobs, but it is the whole appearance. It looks like those weird pseudo grotesque hentai characters, strange.
>>108863144
I am going to try this at least until I can get a more normal body type. The whole character looks overly bimbofied, It's weird because I never had this issue with previous base Illustrious, Noob or NAI.

File: burgerhallucination.png (559 KB, 896x1152)

559 KB PNG

>>108863171
feed your hallucinated girls some burgers bitch.

>Want tan tomboy complexion
>Add Tan
>Adds tanlines aswell
>Tan, (tanlines:-1)
>50/50 chance of coming out brown
>tan, (tanlines:-1), (dark skin:-1)

>>108863245
If there are no tanlines, how do you distinguish tan from dark skin though?..

File: 00343-1645255764.png (560 KB, 1600x1280)

560 KB PNG

>>108862881
mogu mogu
less angrily though and more in an act of defiance

File: Anima_00051_.png (1.05 MB, 1024x1024)

1.05 MB PNG

>>108863276
Thanks man, I've been trying but I can't get it lol Teto just keeps shoving it into Miku's mouth (Which works for the anger part), but I wanted her to bite it via angry or annoyed protest.

>>108863274
>how do you distinguish tan from dark skin though?
How do you distinguish red from blue you fucking idiot?
What sort of dumb fucking question is that? Are you some monkey esl sopa de macaco that's tries to drag southern europeans to the same subhuman level as you?

>>108863307
>red from blue
More like brown from brown KEEK

inside you there are two /adt/'s
devil:
>>108863307
may cry:
>>108863274

>>108863304
yeah I kept getting the same. I only managed that with a lot of rng and then spent 15 minutes manually cleaning it since it was the best I was getting.
<https://files.catbox.moe/my3arl.png>

also getting really annoyed at how baked in miku's headset is. have to put headset in negs and "no headsets" in prompt to even get it to be removed SOME of the time.
also this would be a whole lot easier if we had actual controlnets with anima. Upscaling on it is really not great(while genning higher res at base t2i is fine) and it would've been easier/faster to iterate through lowres and try and modify that to upscale.

File: Anima_00052_.png (1.01 MB, 1024x1024)

1.01 MB PNG

>>108863341
I'm just a shitter using Comfy's template but it works some of the time unless I'm extremely specific.
https://files.catbox.moe/4lofft.png

>>108863351
>パンツ
??

>>108863341
A lot of things are super baked in sadly. Some characters have multiple tagged costumes and so you can prompt a specific one but still get parts from the default costume. Sometimes even the other way around where you prompt default but get aspects of another one. Same if you prompt custom clothes, it might force stuff like neckwear from default design.

>>108863362
it means family.

>>108863362
It was randomly put there.I have no idea honestly why it put "pantsu" there.

File: 0.png (1.48 MB, 1344x1728)

1.48 MB PNG

>>108863245
You're fighting against yourself, try "tomboy, (dark skin:0.9)" and adjust down instead of trying to jam into a super narrow space

>>108863446
now this is something

>>108862814
>it can't lock down text
It absolutely can.
Turn up the Canny/Anytest strength and/or turn down the denoise. Not complicated.

>>108862814
>>108863535
Also, for text, you're much better off adding it with an image editor then inpainting over it with high Controlnet strength and low denoise than you are trying to "lock it down" in an initial gen. If you're having problems "locking it down" in your initial gen then yeah, I get why you're having problems. Don't do that. Just manually add it with an image editor then inpaint over it. Using Controlnet when you inpaint will preserve the text with a higher denoise than you'd be able to without Controlnet. That's the primary use case for Controlnet and text.

File: :D.png (19 KB, 168x144)

19 KB PNG

>>108863446
computer enhance the image. zoom in.
oh my god. it's him.

File: Anima_00060_.png (779 KB, 1024x768)

779 KB PNG

>>108863934
That's really fuckin' good! I'm still trying with random styles. (You can see by the signature I keep telling it to not have.)

File: Anima_00062_.png (881 KB, 1024x768)

881 KB PNG

It recognizes Kurenai Yuuji?! HOW?! I don't have a LoRA for it!

>>108864077
Perfect. "Just accept it."

File: Anima_00066_.png (801 KB, 1024x768)

801 KB PNG

Watch the teeth Teto! You could take a finger off doin' that!

>page 8
Oh nyo...

File: Anima_00068_.png (829 KB, 1024x768)

829 KB PNG

lol sorry I was playing a bunch of Payday 2.

File: ComfyUI_35000_.jpg (607 KB, 1728x1344)

607 KB JPG

>>108864782
what about your dayjob

>>108865006
https://www.youtube.com/watch?v=sCHkdFO8oCo

>>108863446
If there were a chatbot of this, I would appear in the scene, kill the villain, and do savior fagging/slowburn/romance with the Blue Archive character.

File: 00011-424061505999.png (880 KB, 1024x1024)

880 KB PNG

goodbye forever comfyui. I won't miss you.
neo has a long and bigger dick with foreskin unlike you btw ;)

>>108865216
B-based!?

>>108865216
Pretty based,but I'll still use it because I'm a lazy faggot who doesn't want to change.

uncut dicks are just superior
can't blame a nigga for giving in

>>108865216
How is Neo nowadays when it comes to memory management? I stopped using it in February because if I loaded another checkpoint or model, it would lag a lot.

>>108865262
tru :)
>>108865307
i just got an out of memory error with Anima when genning at 1536x1536. am i cooked chat?

>>108865413
Shit, okay staying with Comfy thren

>>108865413
do you only have 8GB of VRAM or something? because that's only taking 9~GB for me and I don't have vram flags on.

>>108862786
full KNOT

File: ComfyUI_35060_.png (1.45 MB, 1536x1536)

1.45 MB PNG

>>108865413
comfyui can do that with just 4G vram on my machine

>put topless in prompt
>bikini still there
>put topless female in prompt
>it refuses to remove the bikini
actually useless trash, I am quitting AI and learning to draw

>>108865621
did you try to prompt "nipples"

>>108865413
How? I can gen at 1536x1536 with my 1660ti 6gb

Chenkin Anima when?!?!? I don't want my anime model to know Peter Griffin or Bugs Buunny or @stan_lee and @todd_mcfarlane. It's awful. It's the equivalent of or worse than e621 or furry dataset!

>>108865870
anon you might wanna sit down for this.
https://danbooru.donmai.us/posts?tags=peter_griffin

If only there is a model that is SDXL size with the quality and anatomy of Flux

>>108866039
ever hear about inpaint?

File: 00024-3792105265.png (2.44 MB, 1536x1536)

2.44 MB PNG

>>108865498
>>108865868
it worked after disabling sage attention.
>>108865568
cute

>>108865925
Poisoned dataset!

>>108865621
Are you slopping it up with score_ and other quality tags? Those are always the main reason anything goes wrong for me. Same with Illustrious, I never use the damn things. They can improve the aesthetics sometimes but most of the time it's a net negative and kills model freedom to follow prompts properly.

I have a question. I understand SDXL only uses DanBooru vocabulary, and anything outside it likely won't be understood. For Anima, which has a more modern text encoder, can I use "vibe tagging" for dataset captioning with Gemma 4 instead of the outdated 2024 WDTagger to tag beyond the DanBooru vocabulary?

>>108866649
Probably, but from what I understand when not using booru tags it's better to use descriptive sentences instead of just comma seperated words found in the image. However it might work fine, the only way to know is to try I guess.

File: kohya-ss_Anima-LLLite at main.png (11 KB, 1565x49)

11 KB PNG

>We have released the ControlNet-LLLite model in the style of anytest, trained on Anima-Base v1.0: anima-lllite-any-test-like-v2.safetensors. It has been trained for a total of about 12 epochs. It has become more effective than the 3-epoch beta version.
https://huggingface.co/kohya-ss/Anima-LLLite
Anytest for Anima ControlNet LLLite, thoughts?

Anyone want to see if this training could be used to create a vibe transfer or artist_collaboration lora for the class?
https://civitai.red/models/2587631/nude-filter-or-simple-edit
I would do it but I know if it works I wont remember to upload it anywhere later.

>>108866932
lllite is garbage.
wake me up when actual controlnet modules exist.

>>108866932
Will test it the older ControlNets were bad.

>>108859807
Shit OP and garbage changes to OP

>>108867054
based

>>108867054
>>108867068
Don't care, 2huai it's safe now.

>>108867140
based

>>108867140
Wait *these* faggots are responsible for this?

>>108867419
Hey, careful with the anti semitism, buddy.

File: 4cute.jpg (1.83 MB, 1650x2200)

1.83 MB JPG

>>108867705

>>108859807
Can you make consistent manga for free yet?

File: file.png (684 KB, 824x641)

684 KB PNG

>>108859807
Minor: Her eye is transparent.

>>108868359
nega teto

File: saba-owie.png (1.73 MB, 1536x1536)

1.73 MB PNG

>>108867829

>>108868359
No worries, I don't feel guilty about it either. I recognize that I am not someone with the ability for this.

File: Untitled.png (4 KB, 471x145)

4 KB PNG

Weird, why does Forge Neo have a Shift value for Anima but ComfyUI's default workflow doesn't? What node does the same thing? I think Shift values control where the scheduler puts emphasis or something.

>>108869299
ModelSamplingAuraFlow

>>108868836
who did this to her

File: adt.png (1.8 MB, 1728x1344)

1.8 MB PNG

>>108869007
>I am not someone with the ability for this
Editing? The fuzzy select tool in most modern editors makes quick work of it, no skill needed outside of setting a good threshold. I did this with krita because I don't really care for gimp and I like that you can apply a grow value to expand the selection.

You could also just not remove the background, it looked fine as white.

File: logo.png (666 KB, 2453x806)

666 KB PNG

>>108869422
also here's the logo with the leading slash

>>108869320
ty

Thanks to the anon who recommended PC Prompt Schedule for ComfyUI https://github.com/asagi4/comfyui-prompt-control yesterday, indispensable for Anima, the best way to control artist tag weights, stuff like [ (empty): @artist: 0.5~1] and all that.

>>108869974
do you have an example workflow with it? and did you save any tests? i've seen people mention this before but i don't know how to use it

>>108870006
The example workflows come with the custom nodes, I downloaded them, went to Templates, then Custom Templates on the left, and selected the PC nodes.
There are 2 options there, can't remember the names but it's obvious which one when you see it, then I messed around with them for a while before adding them to my workflow. Comfy really needs to add a text encoder with these features to the core already!

File: yankee-drawtubers.png (1.4 MB, 1536x1536)

1.4 MB PNG

>>108869330
picrel
>>108869974
np
>>108870006
You can just use PCLazyTextEncode in place of the normal CLIP text encode node. If you want lora scheduling as well, you need to break the prompt out into a multiline string node so you can feed it into both the text encode node and PCLazyLoraLoader. As anon said, just look at the t2i template from the node pack.

Anima feels unstable desu, I have a hard time keeping the style consistent between different seeds and resolutions. I probably need to train a lora at least for the style so it becomes more deterministic at least for what I want the gen to look. "Newest" gives me this seasonal anime type coloring, while "year 2025, year 2024, year 2023" give me more of that VNish coloring instead of cel shading.

>Anima can do 1440x2160 denoise 1 natively without artifacting
Guys... I'm gonna need more lube

Then there's the face too, I already knew how to draw the character's face, so I thought "Anima will keep it consistent between seeds&tags, easy peasy" but I was wrong, even with Euler Normal it's a pain in the ass, swap to Euler A or Er SDE and the entire tag universe reinterprets itself and I have to start over again.

>>108867829
cuteboxes?

>>108870562
it's ebifurya and ixy

File: 450x_auto__so_(1).jpg (121 KB, 450x788)

121 KB JPG

>>108870441
>>108870467
You need a shitmix ASAP, and no, in your emergency case WAI Anima is not enough. You need a pure shitmix, https://civitai.red/models/2606801/nova-orange-am?modelVersionId=2962938!

>>108870614
nta but i downloaded it
it is as terrible as advertised thanks

File: ComfyUI_00047_.png (3.59 MB, 1440x2160)

3.59 MB PNG

Due to natural language, it will handle multiple subjects well. One of my tests is to have one girl with white hair and a black bikini and one with black hair and a white bikini. It passes the test easily. Illustrious/Noob would get it wrong at least 30-50% of the time. It does not warp proportions when generating natively at a fairly high res. I haven't gone all the way up to 4MP though. It specifically does not warp the heights (at least nearly as bad) of age difference subjects to have their faces line up, and does not usually seem to make dicks comically long to try to get them onto faces or between breasts. Hands are extremely fucked and backgrounds are pretty messy. First time I've had fun collecting LoRAs since 2021. picrel uses only detail LoRAs without artist tags so as not to artificially fix any stability problems.

File: tengno.png (313 KB, 1490x1135)

313 KB PNG

made lora from these https://x.com/tengnose1988/status/1547788119960719360/photo/1

>>108870698
this reminds me of nyxa...

>>108870599
b-but the rest of the settings

File: 1519771671892.jpg (31 KB, 625x480)

31 KB JPG

>>108870707
Can you upload it?

>>108870707
Seconding the lora request and I would also like a catbox for the second from top right because I love that pose and sass and don't know how to prompt it.

Also I wanted to get a dataset of images from twitter but I have no idea how to mass download galleries anymore, everything seems so complicated, nothing worked so far. Not that I can make a lora, but I was going to make a dataset kek.

>>108871277
>get a dataset of images from twitter
gallery-dl. You will also need to get a cookies.txt from your browser to scrape twitter with this

>>108871446
A cute!
Now take her clothes off!

>>108871277
>>108871408
Sure, basic prompt is
@tengnosetypeb, pixel art, standing, full body, looking ahead
negative: worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, sepia
Add thick thighs to negative if you dont want them coming out too curvy
gen at 832x1216, to turn it into its true pixel size resize to 139x203 (or 16.66%) nearest neighbor
https://files.catbox.moe/443aha.safetensors
https://files.catbox.moe/exxw3d.png

File: 00144-342319931-a84286df-(...).png (2.19 MB, 1728x1344)

2.19 MB PNG

>>108871453

>>108871470
What a bitch!
Rape it is!

>>108871466
based ty

File: 00148-342319931-e85102f5-(...).png (2.22 MB, 1728x1344)

2.22 MB PNG

>>108871478

>>108871492
But it gotta wait, I just got to work.

And I've never did a rape gen, so I might struggle getting the pose and feel right. And I gotta replicate the dress and style.
But I like a challenge!

>>108871607
Cute but very fucked up depth of field.

File: 00192-1334012249-da5fa49f(...).png (2.98 MB, 1344x1728)

2.98 MB PNG

>>108871621
I feel like my efforts to actively make it more fucked backfired.

>want to generate anime slop
>don't want to spend several hundred dollars more on a nvidia GPU that's nearly-identical in performance to its AMD equivalent for everything that isn't making slop
pain

>>108872233
you can spend hundreds of dollars on an online slop service instead
isn't that great?

>>108871594
Chibi~ :3, can't wait for AnimaYume base, sisters!

>>108872233
Can't you gen on AMD?

>>108872638
You can. It's quite slow compared to Nvidia at MSRP. You're not finding Nvidia cards at MSRP.
AFAIK it Just Werks™(on Linux). This is assuming Anon's on RDNA3. If he has an older AMD card, he's fucked.

File: image.png (43 KB, 647x463)

43 KB PNG

>>108870752
these settings were revealed to me in a dream

>>108872770
As a Linux user with RDNA2 I can confirm none of the shit works at all.

>>108872866
Higher end RDNA2 technically works from what I've read, but ROCm performance is a joke on there, and it's beaten by even low end RDNA3.
I'm running Ampere on Arch so I can't attest to anything personally.

>>108872839
>revealed to me in a dream
Based
>0.55 denoise
Slopper

At what point does denoise stop being fixing stuff or adding detail, and cross the slop threshold? For me it's 0.25.

>>108872965
It depends on sampler and cfg. For example higher cfg changes image more with the same amount of denoise. You always have to think about these things in combination.

>>108872965
Don't know if anyone was able to solve the mystery, but the latent upscaler with values less than 0.5 leaves the image distorted. That's why the anon in the image >>108872839
wanted to, or rather was forced to, use 0.55 denoise because he was doing latent upscaler.

>>108872233
>>108872638
>>108872770
im on rnda 2, with forge neo (and windows) i can at least gen illu and anima quite decently
havent tried video models, last time i tried they were a huge pain
training loras is a PITA too, still use rented nvidia for that (lol 48gb vram cards)

>>108872965
Depends on the scheduler. kl optimal, AYS, GITS are pretty good at preserving original geometry even at high denoise ratios.

>>108871446
>>108871470
>>108871492
Bitch deserves rape! >>>/h/8884660

>>108873019
nta but everyone knows (or should know) that latent upscaling has always required more denoising
this guy >>108872935 is just a moron

>>108873507
teto really let herself go

>>108873945
it happens
thankfully hags are in demand

made my first dataset and trained my first lora
picking the right epoch before overtraining occurs isnt that easy though, guess Ill just use a slight overtrained one and reduce lora weight

File: heh.png (189 KB, 704x530)

189 KB PNG

>>108866019