File: bt5g3n.png (2.25 MB, 1024x1536)

2.25 MB PNG

SD1.5 throwback

File: file.jpg (806 KB, 4096x2048)

806 KB JPG

its getting there

blessed thread of frenship

>>110275517
nice, moss?

>>110275534
yes but its a small test run specifically on this: https://huggingface.co/KBlueLeaf/MOSS-Image-heal

gonna bake one for regular moss, kromoss and then compare all three

>>110275552
>https://huggingface.co/KBlueLeaf/MOSS-Image-heal
This looks like crazy placebo.

>>110275566
its supposed to be "easier" to train, which im testing, but yeah for regular generation i just stick to kromoss

File: 53cea2078626354e21237c5f6(...).jpg (72 KB, 1024x720)

72 KB JPG

>>110275614
>we're all here

File: emma blows a kiss.png (784 KB, 880x1184)

784 KB PNG

The anon who suggested to mix resolutions in training seems to have a point.
I ran the exact same parameters, seed but did evenly mixed 768p / 1024p instead of full 1024p (He suggested 512p/768p/1024p but I didn't feel like training on 512p) and the result is better. Haven't run into body horror.

>>110275683
Those arms belong to a man

>>110275706
absolutely not

>>110275614
suffer, worm

>>110275706
That's a natural athletic 1girl.

>>110275683
>doesn't even post metadata

>>110275706
and that man is me

I just set up ComfyUI for the first time, how common is it for image models to be hard or soft censored?

I started with Flux 3 Klein and the results were atrocious. Then I tried Krea 2 Turbo and people actually looked like humans

>>110275765
every single released model is incapable of making dicks and vaginas (exception beign moss), they all require a community trained version for porn

File: ComfyUI_temp_ygliz_00069_.png (2.83 MB, 1165x1747)

2.83 MB PNG

>>110275738
U want the catbox?

>>110275765
Flux models: hardcore safety-maxxed, not worth your time.

Krea2 models: soft safety-maxxed, easily uncensorable models, they are great. I just use normal Krea2 turbo + loras.

Moss models: cunny-maxxed, not censorship at all. Pretty good models but they are small so their anatomical knowledge is inferior, borked hands can be common.

>>110275770
I'm just looking for a general purpose model to supplement other things, such as my existing harness. I'm not doing dedicated genning right now. I'm already used to working around LLM censorship, but I wasn't expecting image model censorship to be worse. For example, vanilla Flux 3 Klein failed every test I tried with humans.

>>110275813
If it's for a little bit of everything, MOSS or Krea 2 Turbo?

an actually good resource for prompting

https://cari.institute/aesthetics

>>110275839
holy fuck this is gold.

>>110275839
90% of those will mean absolutely nothing to QWEN

>>110275854
you can still use your eyes to absorb the style into your brain, and then prompt for it, kind of like training a lora in your brain

>>110275854
>The absolute state of promplets

File: MOSS-er_sde-2.5-25-2026-1(...).png (2.74 MB, 1088x1920)

2.74 MB PNG

hhrrmm interesting stuff

Subject & Pose:
Full body illustration of Princess Peach, dynamic slightly angled 3/4 front view, right hand holding her iconic yellow pink-trimmed parasol high over her shoulder, left hand delicately hovering near her waist. Expression is cheerful and wide-eyed with a bright open-mouth smile.

Character Details & Proportions:
Vintage 1990s character design proportions. Slightly oversized rounded head, expressive bright blue eyes with minimalist single-stroke eyelashes, simple soft arched eyebrows. Voluminous, softly rounded golden-yellow hair styled in low sweeping side-flaps and pointy top bangs. Wearing a cyan-jeweled gold crown with red rubies, a classic bright pink gown with dramatic puffy dark-pink shoulder sleeves, rounded dark-pink waist panniers, and a circular cyan brooch centered on her chest. White evening gloves extending past elbows.

Art Style & Medium Execution:
Authentic 1990s Nintendo promotional ink and watercolor box art style, official art studio print, visual work by Yoichi Kotabe and Shigefumi Hino. Clean, variable-weight bold black brush ink outlines defining all edges. Soft wet-on-wet watercolor cell shading, rich analog color gradients, subtle watercolor paper tooth texture, clean white drop-shadow backlight background. Flat vibrant color palette, bright primary colors, warm nostalgia lighting, crisp clean lines, zero 3D rendering, zero realistic shading, zero CGI textures, classic 2D animation cel aesthetic.

Not sure Google Gemini understood the assignment fully but i'm the type to give full credit anyway.

It's nice to see debo having fun in his little make believe thread.

I consider this the real bake because the other one snubbed my very good gen. Thank you baker!

>>110275829
At this point, I would say go Moss for everything, Krea 2 is a great and fast model, but Moss base model with the dmd2 turbo lora is just better, both for artstyles and realism and realism in particular since it doesn't have the tell-tale 'aesthetic tuning' look that all the distilled models have.

Also it trains incredibly well AND fast since it's a small model, I was expecting Krea 2 to rule until Krea 3 came out, but Moss just came out of nowhere and is incredibly good.

>>110275941
If you make my collages it's because you've made a high effort and or high quality gen.

>>110275880
This happened even before the split, it's crazy how he can do the same thing for years and gets surprised nobody pays attention to him and ignores him.
His nemesis is even missing you would think he would chill the fuck out for one day.

>>110275975
>Moss base model with the dmd2 turbo lora
really? Could you show some examples? I've tried the turbo options on the sft moss model and it makes everything really low detail and cripples a tiny bit of adherence. People were saying NOT to gen on base under any circumstance.

>>110275983
>People were saying NOT to gen on base under any circumstance.
Who has ever said that ? It was clear early on that base was the best, and cemented with dmd turbo lora

>>110275996
>Who has ever said that ?
your momma, anyway grabbing base now.

>>110275996
>It was clear early on that base was the best
I don't know why you keep peddling this blatant misinformation. The devs themselves have said base is not for genning.

>>110275779
Yes please

>>110275829
Krea2: Doesn't has the NSFW knowledge of Moss so you need to add loras, sometimes very specific loras. Fantastic prompt understanding, you can easily iterate your prompt until you get what you want. Very good anatomical knowledge, borked hands are rare.

Moss: You don't need loras to gen a lot of NSFW concepts. Inferior prompt understanding, borked hands can be common.

So, depends on what you want to spend your time, hunting and balancing loras or re-rolling the dice for good anatomy and the specific image you have on mind.

https://strawpoll.com/GeZARJ3wjyV

>>110276022
sft is recommended for drooling retards who expect the model to output slop with minimal prompting, even when krea2 was launched power users figured out quickly that using the base model with a turbo lora loaded in at a lower strength was the best move

File: moss-t2i-base-turbo-lora-(...).png (1.89 MB, 1280x1280)

1.89 MB PNG

>>110276022
You are retarded, people were preferring Base over SFT early on, and when dmd turbo lora dropped, nobody looked back.

Are you seriously using Moss SFT for genning ? Are you daft ?

>>110276069
damn, what a high quality gen that SFT could definitely never do....

>>110276058
>>110276069
You could at least try to be subtle

>>110276023
catbox is being a shit again
https://h.uguu.se/LJKLqzCC.png

>>110276093
>get blown the fuck out
>accuse the person of trolling
I humbly accept your concession.

File: moss-t2i-base-turbo-lora-(...).png (2.2 MB, 1280x1280)

2.2 MB PNG

>>110276090
It's prompted to be a vintage magazine scan and the first SFW images I found in my output, show me your best SFT, I'm waiting

>>110276022
>The devs themselves have said
I frequently disagree with model devs. Anyone who gens a lot does.

>>110276093
Not samefags, try again

>>110276116
nta but these
>>110275779
>>110275683
are sft

So is there a meta for Moss captioning? A simple natural language paragraph seems to work fine, at least for low rank character loras, do you get any noticeable benefit by bbox autism or anything else?

>>110276154
Asian chick looks really AI sloppy to me, but the other looks nice, give me prompt and I'll run it through base

>>110276183
>>110276103

>>110276177
From what I've heard on training discord channels, bbox training is essentially a waste of time, Moss works great with natural language, for training, and also tags from what I gather, but I doubt people are interested in going back to tag-based prompting.

>>110276212
>>110276177
So did both of you just ignore the skills file that basically tells the LLM what to do?

File: sftt.png (3.42 MB, 1434x1434)

3.42 MB PNG

>>110276116

>>110276212
Yeah sound about what I expected. As for tags, I have a weird experiment training run using old dataset with tag captions for noobAI going on in the background currently. I don't intend to use tag captions for anything serious, but I want to see how it responds to gauge how flexible it is. I will see if I can corroborate claims.

Blessed thread of frenship

>>110276234
What important info relevant to the conversation is here?
https://github.com/OpenMOSS/MOSS-Image/blob/main/skills/moss-image-prompt-enhancer/SKILL.md

>>110276264
That is just plastic slop anon

>>110276350
What's this artist?

>>110276340
Sorry, meant for >>110276116

What should I generate?

>>110276340
That's just a regular gen without the colors being crushed by a turbo lora

File: anima_aestheticV11_00859_.png (882 KB, 1232x1584)

882 KB PNG

>>110276361
>nigga doesn't know
>*smacks lips*

>>110276350
>>110276417
no we don't know your dime a dozen sheboon enjoying artist.
nice stuff though.

File: mybad.png (2.87 MB, 1434x1434)

2.87 MB PNG

>>110276340

File: anima_aestheticV11_00860_.png (1.23 MB, 1232x1584)

1.23 MB PNG

>ayo the fuck u lookin at nigga
>hurry up and tie my shoes

File: MOSS-er_sde-2.5-25-2026-1(...).png (2.46 MB, 1088x1920)

2.46 MB PNG

>>110276430
what dem boon lips do doe

File: anima_aestheticV11_00863_.png (1.22 MB, 1232x1584)

1.22 MB PNG

its namespace

>>110276154
>these sre sft
They look it.

All of genning is basically divided between people who think SFT, turbo models, etc. look real enough, and those who don't. Maybe this is related to development of some part of the brain, who knows. The former group cannot understand why the latter choose to struggle with less-compliant base models: bad hands, visual artefacts, blur, etc. The latter cannot understand why the former are complacent and accepting of slopped-out gens whose evident fakeness is the first impression made on the eye.

This division in the community has been obvious since the earliest SD1.4 finetunes, and it's not going away any time soon.

File: 1785669644024665.jpg (1.18 MB, 1552x1936)

1.18 MB JPG

>>110275439
https://files.catbox.moe/mgbv7r.png

>>110276501
Honestly for me i've genuinely just been out of the loop for a while and missed the boat on the community discourse for MOSS.
SFT has its issues, undeniably. But fellas in this general are unironically fine with mushy fingers/hands for some fucking reason.
Like in this example >>110276443
rarely i'll get extra fingers with the INT8 quant of SFT. Which is why after my lora finishes training im gonna try shifting to Base and see how that goes.

>>110276501
meds

>>110276501
>and it's not going away any time soon.
:(

>>110276501
Also, since we're comparing gen realism, my gymnast 1girl in the OP (bottom-right) is moss base, ~1.9 cfg iirc.

>>110276477
Thanks

>>110276506
Will you upload lora to somewhere else?
Litterbox was working snail speed for me.

File: moss-t2i-base-turbo-lora-(...).png (2.25 MB, 1280x1536)

2.25 MB PNG

>>110276191
Just came back and every queued gen had nipple slips, so the kromoss turbo lora really don't like covering up areolas, so I slapped on a tube top

File: bwa.jpg (4 KB, 226x223)

4 KB JPG

>>110276569

>>110276569
>kromoss turbo lora
i know to avoid that one now, that looks absolutely dogshit.

>>110276526
>my gymnast 1girl in the OP (bottom-right) is moss base, ~1.9 cfg iirc
cool thanks for the info actually, was gonna ask for the catbox last night but didn't.

>>110276548
https://www.mediafire.com/file/jymmvtdovw9p2ek/shejanto-step00013178-6de743-state.safetensors/file

File: moss-t2i-base-turbo-lora-(...).png (2.13 MB, 1280x1536)

2.13 MB PNG

>>110276580
>>110276575
You have no taste, it's sad really, fapping to slop like pajeets

>>110276598
It's ok anon, Thanks for sharing. Others will be able to form their own opinions.

>>110276569
nice stubble
>>110276598
based

>>110276511
>fellas in this general are unironically fine with mushy fingers/hands for some fucking reason.
And "fellas in this general" are unironically fine with slop-lit easy-to-spot AI images etc.

Like I said, the two sides can't understand each other. Some scrutinize details and sperg about hands, others are very affected by the image taken as a whole and sperg about "plastic" appearance/etc. This is exactly the divide I am pointing to. The latter would call bad hands a trade-off, the former don't see the gains that could justify it. This exact argument has been happening on and off since 2022 at least.

I keep images with bad hands all the time (if the mutant 1girl is cute enough), but I avoid posting them. I try to only post my most compelling and blemish-free gens

>>110276569
It has that turbo texture, reminds me of z-image

Literally don't care. I'll keep posting my kinos and there's nothing you can do to stop me.

>>110276501
>>110276626
Trvth

File: 2026-11-08-165220_937x541(...).png (51 KB, 937x541)

51 KB PNG

https://strawpoll.com/GeZARJ3wjyV

>>110276598
>no taste
nigger wipe the cum off your monitor the other one objectively looked awful, like flux 1 awful.
this one actually looks totally fine, no more plasticyness, just a natural imperfect grainy look.
>>110276626
I mean that's how i see it but i'm not completely lobotomized so i can see the nuances here. An extra finger here and there doesn't completely kill an image with overall good composition (also there's inpainting/editing it out so)

you know moss won when anon is arguing over which version is best

>>110276657
real.

>>110276657
I'm surprised the schizo hasn't rehashed the whole "catastrophic forgetting" troll since, like Anima, Moss suggests freezing the encoder during training.

>>110276589
Thanks

>>110276646
why slopft

>>110276598
>>110276569
am i going crazy? these gens look like shit

>>110276745
You're not crazy anon.

>>110275983
>People were saying NOT to gen on base under any circumstance.
those people fell for the noob trap kek

>>110276745
naw it's just a reminder most of the interactions in these threads are bots or crazies, the second the word schizo pops up you know we're outnumbered kek. I forget that when i'm not lurking as much as i used to.

>>110276684
Schizo melted down over anima and came up with many lies because it got comfy funding and he endlessly seethes about comfy.
Some of the lies also got coincidentally signal boosted by discords with competing products to anima.
He doesn't really have any strong reasons to seethe about moss, at least to a similar deranged degree.
>>110276745
The grainier the gooder, the realer and the kinoer.
Learn to gen chud.

>>110276598
>>110276569
Reminder that these are the people telling you not to use SFT

>>110275983
>People were saying NOT to gen on base under any circumstance.
Something you learn here pretty quickly is that "people" do not know what they're talking about. You must free yourself from what "people are saying".

>>110276745
No, you are just dealing with someone who is either a troll or, more likely, a deranged schizo

>>110276782
many such cases

File: moss-t2i-base-turbo-lora_(...).png (776 KB, 768x768)

776 KB PNG

>>110276745
Your brain has been fried by slop, probably used Flux for too long

>>110276801
this is like if z image turbo had down syndrome kek

>>110276801
your gens are trash

>>110276382
your mom as a sexy loli
or your dad in his 20s as a woman

>>110276501
I think you are into something. Other things may be related.

I don't think there is anything with wrong with SFT Moss, but I prefer Krea2 look. I like the faces of "perfect" Asian models, the ones that some people refer to "plastic" looking. CGI in anime doesn't bother me at all unless is really bad. The uncanny valley produce nothing on me, I recognize it but it doesn't alter me.

File: moss-t2i-base-turbo-lora-(...).png (2.1 MB, 1280x1280)

2.1 MB PNG

>>110276801
imgimg test, 640->1280, ok but not as good as straight up 1280

>>110276816
everyones gens here are trash
most worthless general

>>110276809
>>110276816

>>110276839
imgimg test, retards

>>110276846
You can't say that when >>110276881
exists

>>110276833
I don't want to do either of those.

https://x.com/ComfyUI/status/2119528573539320147

>Count down starts ...
>And no, it's not fundraising announcement this time.

what is it

>>110275983
Even the MOSS devs tell people not to gen from base. You might get an occasional good gen from base, but that's out of dozens or hundreds of gens.

>>110277000
Probably some real dumb shit, considering it's cumfart...

File: debo_ic_km_00049_.png (2.77 MB, 1792x1076)

2.77 MB PNG

>>110277000
nodes v3
all nodes run on the cloud
pre-order for a discount on tokens

>>110277000
New video model ? Krea 3 ? Or something specific to Comfy (then it is 99% sure to be shit)

>>110277117
we don't need a new video model. h3 is honestly perfect. its probably something not worth talking about. yoland loves dropping vague tweets to hype up dumb shit

>>110277100
Thank you for your valuable contribution.

File: 1786127713368231.jpg (983 KB, 1552x1936)

983 KB JPG

>>110276908
thanks anon

>>110276908
kill yourself

The other thread died without all the samefaging, looooooooooooooooool.


>>110277270
What went wrong on your life to make you so bitter, anon?

File: FVCT.gif (22 KB, 220x221)

22 KB GIF

>>110277303
FACT!

hm...
what to train what to train...

Ace Step 1.5 XL Base here to minister to souls:
https://files.catbox.moe/n5a4fv.wav
https://files.catbox.moe/8wj3t0.wav

>>110277174
The little dithering is really what makes it for me. are you adding it in post? No hate if you are.

>>110277174
yeah this style and all your gens are cool as fuck dude i wish more people posted stuff this unique.

*I* posted industrial cello, that's not run-of-the-mill.

>>110277610
t2i with the lora no post
>>110277621
im gunna retag and retrain my metalheart dataset for moss and see how that looks

>>110277117
could be that new acestep model

>>110277639
>t2i with the lora no post
That's are really big W for moss.

File: ythutghuyt.png (502 KB, 1320x1315)

502 KB PNG

the fucking ASSES and PUSSIES MOSS can put out are fucking INSANE, even with my character lora only having maybe a few images of the subject with her pussy out MOSS just adapts it so amazingly well, Krea 2 and ZImageTardbo never could compare even halfway. Really pays to just have a base model that isn't raped in some capacity like they are.
Honestly, not too upset i missed the initial hype wave as there really isn't much to it, the model doesn't seem to need many sliders or loras to do anything unlike krea 2 and others i used extensively in the past.

>>110277700
i doubt the veracity of your statements.

>>110277711
I doubt the veracity of your intelligence.

>>110277711
I find the veracity in your lack of faith disturbing.

>>110277700
kroma 3.1 pusses still look like open wounds especially if you prompt for spreading or a closeup

I doubt the felicity of your chat logs.

moss gens the best assholes

>>110277711
have you not tried moss yet

>>110277000
It's a Suno v5-tier slopped music model ("Ace Music") >>110270530. Not even joking. I don't why if they're using synthetic data, they don't just use data that has sovl. None of them have yet caught up to Udio in terms of out of the box sovl https://www.udio.com/songs/pHLn2TLb2jSsQfa9CC7S7J
https://www.udio.com/songs/heDdLtrK4wJcbQxShJW7sG

>>110277739
>udio

you're out of your mind.

None of the other models can do this:
>>110277583

moss generates the best feces

>>110277174
You've convinced to bake my first lora. any recommended reading? software?

>>110277583
>>110277773
is this a joke

Moss is a great step forward, I hope video models follow suit.

genAI (in general) is definitely not there yet but my dream is to someday gen a video of a beautiful yogini (that is female yogi) giving birth, where the delivery is completely perfect and effortless due to her mastery of all the muscles of her body/'canal'. Something that leaves you dumbstruck in awe of the potential for the human being to be fully governed over every inch of his/her body by divine Reason, complete harmony of will and matter

moss is only good for realistic stuff, but it knows absolutely no one so you have to train loras that you can't really upload anywhere to make gens you can't really post anywhere

i wish there was a moss only sexy jam

>>110276273
Update: tag captions "work" with moss training. But resultant lora is very awkward to use. Quality is rough too on some gens. Although these are shitty captions from SDXL days, I can't rule out I haven't fucked up something else. So I need to write new captions and train again to speak with more certainty.

>>110277805
i had astra research what was out there and build a slim ui for sd-scripts using its moss_image_train_network.py and networks.lora_generic, set alpha to 1
i caption with gemma4:e4b, bake for 70 epochs (that lora with 300 some imgs is at 13k/24k steps after ~14hr), 1024 with AR buckets only... basically as lazy as possible and it still nails desu

>>110277967
tyvm anon!

>>110277831
I'm not autistic, so I don't like garbage.

is this still the best local song ever made?
https://vocaroo.com/1kzCDVTxuLOk

>>110277863
>moss is only good for realistic stuff, but it knows absolutely no one so you have to train loras
I like making 1girls that look like no specific person, just beautiful

We've already established far beyond the shadow of any doubt that the celeb/deepfake anon is severely mentally ill.

>>110278126
Literally who? Also, I dont understand why people want to generate specific people. No one in real life looks as good as the girls you can prompt for. The added bonus is that they are exclusively yours.

>>110277739
Udio tier is what we would have if Minimax team gave us an audio model trained on the same dataset as H3. It's so sad they refuse to touch real data (from Youtube, which is what Udio was trained on). It's no coincidence that H3 can make short bangers on par with Udio.

>>110278126
>the
There's seven of them.

File: 1666401986577918.png (3.25 MB, 1334x2000)

3.25 MB PNG

>still no flux 3 video
>still trolls here
okay, i back in my h3 cave kek

>Doodle a pose
>Have model img2img it in certain artist artstyle.
>Dont really care about it not being perfect because I just need it for reference.
What local model can I use with a meager 8gb of vram? Comfyui with illustrious has been intriguing to learn with so far but what is the current best thing I can host locally with comfyui? I was fucking around with some subscription models that take natural language and it is a bit scary how good they are at following it.

>>110277863
MOSS responds better to 2D than realism for lora training.

>be chromafag
>enjoy the novelty but like with most ai image models, results lack variation and you begin to see the same face
>shit ton of image models drop since chroma
>still not interested
>discover bigasp3
>never looked back
>even the buttholes are completely different each gen

Despite random body horror and that shitty green tint (which I believe the developer is trying to fix) this model slaps

>>110278423
>never gens hands

>>110278423
>discover bigasp3
stopped reading. cant believe people even still mention that garbage ass model.

>>110278423
Is he still working on it? Last sign of life was 3 months ago.

>>110278309
The model for that is Qwen edit 2.1. I think it can run on 8gb of Vram if you use the small versions of the model and text encoder.

>>110278305
Literally who are you

>>110278466
there was one anon (crazy person) that claimed it was extremely good but every time he showed examples they look like the hottest garbage I've ever seen.

>>110278466
Kek, seethe

>>110278468
Probably, I'm jut judging by his comments on huggingface and civitshite. Also from civits description:

>bigASP 3 is currently still UNDER DEVELOPMENT. This is a very early ALPHA release. Just to give a taste of the model. The current version has tons of problems and will likely not be useful for your everyday gens.

>>110278583
>there was one anon (crazy person)
One of the only people in these threads that has seemed like a genuine basket case and not a faker

He just bumped his thread... a new strategy, a retarded one.

>>110278674
>new
it's not, unfortunately. but you learn to ignore it

>>110278423
>bigASP
Lol, that trash has nothing on either Chroma, Moss or Kroma.

File: 2026-11-08-214120_920x543(...).png (54 KB, 920x543)

54 KB PNG

https://strawpoll.com/GeZARJ3wjyV

How long can your prompts for Krea2 get before things get funky?

>>110278700
Until any of those models reach the variation that I look for, I don't care.

>>110278794
I've had insane 500+ word prompts and it liked it and didn't miss anything.

>>110278423
why yes, I too, am a fan of bigSHIT#3_BimboSlopEdition.safetensors.

>>110276719
It's possible to game polls like that easily

what are your favorite models by category? for instance, edit, realistic, anime/2D, video, etc?

>>110278922
cope

I'm having sol slop me some image captioner UI, it's pretty great. What a time to be alive. It just did a week of work in 20min

>>110278934
>I'm having sol slop me some image captioner UI
what

File: 2026-11-08-222551_1144x10(...).png (95 KB, 1144x1019)

95 KB PNG

>>110278942

does someone have a moss bounding box example they don't mind sharing? idk what i'm doing.
>make character box that covers most of the left side of the frame
>it reliably draws the character on the right
????

>>110278953
what model you use for captioning?

>>110278963
gonna use gemma 12B, but it'll be configurable. just going to use any openai compatible API (i.e llamacpp)

>>110278953
incredibly high levels of kino

>need to rotate a scene by a few degrees
>image edit models butcher the image
>minimax h3 does it without issue

>need to place a character in a scene so it matches the lighting
>image edit models makes the character looked photoshoped in
>minimax h3 does it without issue

Wasnt there suppose to be an h3 image edit model or something, if so, where is it?

File: moss-365174243412835_00001_.png (1.07 MB, 736x992)

1.07 MB PNG

Prompting tip: ask ChatGPT/Claude/etc. to describe the appearance of your setting.

Sometimes you know you want e.g. a street in Boston, but you can't think how to describe it comprehensively. So you ask ChatGPT, "Describe to me the appearance of the houses and the street in downtown Boston somewhere. You don't need to be poetic, I just can't think how to describe it and I want a concise visual description. Pick somewhere unusual, maybe a former working-class neighborhood, not somewhere famous."

It will give something verbose:
>[. . .] The streets are fairly narrow and flat, with patched grey asphalt, concrete sidewalks and wooden utility poles carrying overhead wires. The houses stand close together, mostly two or two-and-a-half storeys tall, with pitched roofs, narrow front porches and wooden steps. Many are nineteenth-century houses now covered in pale vinyl siding—cream, grey, faded blue—with white window frames and occasional projecting bay windows. Small fenced yards separate them from the pavement, and cars are parked along both curbs. Some houses retain elaborate old trim, while others have plainly modern doors, windows and railings.

So you cut to just the useful bits:
>Dark blurry 3AM (close-up:0.75) photo of a cuter-than-average young woman, 18 or 19, in Dorchester \(Boston\) late at night. She's sort of hot, skinny through the midsection and just a little bit thick in all the right places, and her (red velour:0.6) dress shows some cleavage. (Behind, a dense packed neighborhood, wooden utility poles carrying overhead wires. 2.5-story houses, 19th century now covered in vinyl siding, close together with pitched roofs, narrow front porches and wooden steps, some with bay windows. Small fenced yards separate each house from the pavement, and a Nissan Sentra is parked at the curb.:0.5) She's making a sexy kissy face for fun while she grabs and squeezes her own tits as a joke. Her tits are magnificent.

(cfg 1.9, moss base)

>>110278928
>edit
Would probably grab Klein 9b again if I need that.
>realistic
Krea, Moss
>anime/2D
Anima
>video
H3

File: img_tag screenshot.png (1.33 MB, 1267x1597)

1.33 MB PNG

>>110278934
lol I attempted something similar with Claude a while ago, never finished it. Was based on Casey Muratori IMGUI-style GUI as explained by Ryan Fleury (of EPIC Game Tools, current developer of the RAD Debugger), has a pretty fast working file browser (it's actually faster as an image viewer than the image viewer I use normally.. lol...) and was planning to add some tools for bulk-tagging of images, indicators for "missing X tag", etc. The main idea was to have some chip-based drag-and-drop tagging, an autocomplete search field for tag searching, as well as recognizing and separating out a natural-language portion of the caption which can also be edited as a simple text field (for Anima). And of course trying to have every kind of keyboard shortcut one could want so it feels natural and easy to use, fully-featured text fields, etc. And of course zero dependencies

If I actually gave a shit about tagging for LoRA training I might not have abandoned it. I was just trying to prove I was capable of making a LoRA because someone in the thread called me out, I got sidetracked because I found manual tagging annoying with the limited tools available. I guess I can't make a LoRA after all...

File: 2026-11-08-232746_1566x10(...).png (226 KB, 1566x1017)

226 KB PNG

I can already smell the kinos

>>110279241
Sounds like you had something pretty specific in mind. I'm just seeing this as a quick and dirty utility to make captioning as painless as possible.

do bbox's work with sdcpp?

>>110279380
No reason why it wouldn't. You may want a separate tool to write them though.

Is krea2 still the current meta or is there another model release dominating shit now

File: img_tag screenshot 2.png (27 KB, 1271x795)

27 KB PNG

>>110279241
There was a lot of manual tidying up still to do of all the relations between tags and their 'groups', based on scraped wiki html files. I had a separate script for preparing the tag data the program uses, this would take a combination of user-written hierarchies, a sorted list of all tags, and the cleaned-up product of the scraped wiki html and produce something ready for the program to use, so the GUI and the data preparation were separate programs.

And there were features yet to be implemented, and features whose shape I couldn't decide on. What would tagging a dataset of hundreds of images actually entail doing? Ideally you'd want to empower the user to write their own checks, e.g., filter to all images with 1girl and highlight in red any that do not include "*boy", "solo", or a host of other tags of which at least one should be present (these kinds of categories would probably have to be defined by me ahead of time for convenience, as a further manual elaboration of the dataset, but maybe also the user should be able to create their own). Or show all images and display tags under them, filtered to only these categories: scene, landscape, background, pattern, etc. So you scroll through and eyeball that every image seems to have the correct tags about its background. etc.

Basically this would require some kind of user scripting language for queries in the file explorer part of the program. And then there's the question of how to do that sort of thing fast if you have e.g. 10,000+ images. Browsing a folder of 10,000+ images is already quick, it just shows a skeleton immediately (in case you're just navigating through) and builds the thumbnail cache in a background thread.. but anyway I don't know what difficulties lie in wait just down the road

>>110279351
There is no limit to "as painless as possible". You can always make it even less painless. "Le mieux est l'ennemi du bien." ...

File: Tionesta-ac-moss2.jpg (3.11 MB, 3072x2304)

3.11 MB JPG

>>110279432

>>110276646
hope you retards know that the sft model produces vertical banding artifacts
see: https://huggingface.co/Kijai/MOSS-image_experimental

>>110279433
Also ofc any task involving interpreting the tag dataset directly couldn't be done with assistance of any of the frontier LLMs, obviously, because encountering Danbooru tags makes them freak out and shut it down. (I had very strict partitions of the script from its own data, and did not allow Claude to look at the data ever.) So if I wanted to automate the task of creating the hierarchies of categories (which should actually be so easy to automate) I'd have to figure out how to run a local LLM that is less scrupulous. Probably not hard but I am so, so lazy

>>110279453
If you used SFT you would know that the banding happens very rarely and it barely noticeable.

>>110278423
You're obviously the one working on bigasp3. If it was any good, there would be more people talking about it. The first two versions sucked. I can't imagine this one being any better.

>>110279457
>If you used SFT you would know that the banding happens very rarely and it barely noticeable.
i did use sft and started noticing it, so i switched to base+sft lora at 0.5 str. but keep coping.

just wanna say im literally living the dream and making the porn of my dreams.

>>110279495
are you saying the banding is present in every single gen?

>>110279496
based and same

>>110279522
moss base?

>>110279527
Yup. Same prompt/settings as >>110279117

>>110279529
Do you think her face looks normal?

>>110279532
Yes, I do.

>>110279535
virgin?

>>110279539
Who, her? Doubt it.

File: forest2.jpg (3.77 MB, 5440x3072)

3.77 MB JPG

Native resolution.

>>110279583
I don't see a single native

>>110279546
cheeky

File: MOSS-er_sde-1.0-12-2026-1(...).png (3.39 MB, 1344x2368)

3.39 MB PNG

zoo wee mama im fuckin lovin' moss so much

Hello anons, sorry for the retardation I'm new to this, the last time I did image-gen was pretty casually on a 3080 with regular stable diffusion in 2022. Or something like that I can't remember. If I have a 5090 and want to generate 3DPD porn/fetish images of women what models should I be using? I'm trying with qwen-image-2512-FP8 (I think that's what it is) but the results are disappointing. It could be my prompting as well, I have no idea what I'm doing in comfyui.
>training
Just now finding out about this. Can you also train audio and video?
>>110277967
Need to figure out what this means, thanks anon
>>110275683
>>110276569
>>110276598
See how do you get kino like this?
>>110275756
Based

Apparently I need to be using MOSS or krea2, and perhaps use my own LORAs.
Also it's not that important to me, but can any local models do text in images along with realistic looking people?

>65/2000 [07:50<3:53:31, 7.24s/it, avr_loss=0.337]
doing it!

File: Screenshot 2026-11-09 003249.jpg (1.07 MB, 3547x1933)

1.07 MB JPG

>>110279241
>>110279310
do you actually spend time reviewing the caption results of the selected analyzed images used in your vibe coded captioner program?

how much do you care about the accuracy captions and precise description wording within the results?

just curious how invested other people are in the captioning process.

lately I've been avoiding downloading character loras that lack decent enough captioning descriptions and tags necessary to generate images that portrays the character's resemblance accurately to their source og material.

>>110279675
>do you actually spend time reviewing the caption results of the selected analyzed images used in your vibe coded captioner program?
Can't speak for the other anon, but my program was 100% for manual tagging and manually fixing tags, it's not a wrapper for an autocaptioner. So if I was autocaptioning I would do that first and then use my program to look at the results and edit them myself.

Making that something you can do inside the program was one of those features I thought maybe I'd add someday or maybe not, but I was more interested in trying to figure out how to make manual booru tagging more ergonomic.

every gen in moss takes four minutes without the lightning lora
why is it so slow? no other model architecture over twice its size is this slow without lightning loras.
could my comfyui just be fucking busted as usual?

File: 2026-11-09-011840_1344x97(...).png (485 KB, 1344x970)

485 KB PNG

>>110279675
>do you actually spend time reviewing the caption results of the selected analyzed images
This is the very first time I do this so I'm going a bit fast just to get the full flow down.

I spent some time tweaking the captioners system prompt so it gives me a caption I'm happy with on a couple test images. Then I ran the batch caption on all images, quickly went over the captions as it was generating to remove any glaring errors, in this case it liked to write "anime-style girl" despite the system prompt being clear not to define stylistic attributes.

I'm assuming style loras are a bit more forgiving on "bad" captions. Anyways, looking forwards to seeing if this will be complete shit or not.

https://files.catbox.moe/esbk7r.mp4

>>110279920
nightwish-esque

>>110279940
Not my favorite band but I'll take it :)

File: moss-894763181920076_00001_.png (988 KB, 736x992)

988 KB PNG

imma need some of my hand experts to get real zoomed in here and figure out whether we're in unpostable mush territory or on the "it's kind of plausible idk" side of the line

>>110280073
certified slop

>>110280073
needs higher base resolution and more steps

File: moss-589071165274845_00001_.png (827 KB, 736x992)

827 KB PNG

>>110280084
I don't think that's true, I think 40 steps is usually plenty (although I'll sometimes do 60) and 0.7 megapixels seems to work fine. Minimum size of training image data was 192x192 according to their report, and trained at original resolutions; and other anons have successfully tried sizes as low as 250ish. I wouldn't go that low just because you're relying too much on the VAE to reconstruct basic image detail—it's not like SDXL where you can gen at 1mp with a conditioning that tells it to create the style of a smaller upscaled image, and then you downscale after. (SDXL was such a cool idea). But I really doubt 0.7mp is going to be problematic that way.

The problem is texting on a phone. There are a lot of ways someone can put their hands when they do that, it's hard to reproduce well as a 2-dimensional visual.

File: Screenshot 2026-11-09 010526.jpg (379 KB, 1431x1690)

379 KB JPG

lots of bad krea2 character loras on civitai especially from certain bakers i once had respect for in the illustrious days. some people are unwilling to admit the inadequacies of their captions. This particular lucoa lora is very bad and broken primarily because of bad captions but also the low training steps and highly plausible low quality dataset used. I spent multiple attempts just trying to get the heterochromia eyes of lucoa generate accurately but had no success. Glad some else did a slightly better lora of her but theirs still feels lacking.

>>110279817
not bad but still needs some work. you should emphasize the swimsuit to avoid confusing the model. "one piece swimsuit" or something specific like "female Japanese one piece student swimsuit". The position of the girl you mention corner will be a problem. The positioning is wrong, it's the left of the image not the right and add "partially out of frame". some vision models fuck up the captions for left and right positioning of the subject because their overtrain with images of people being in central composition. The real concern is that i don't see the emphasis of the visual style. The lack of words to describe the visual style and aesthetic will fuck up the lora. Just a simple artist tag trigger will not be enough especially for modern models. If you have enough energy should popularly emphasize the lighting, highlights and shadows. if you doing style lora, should work and expand on the system prompt of extract get this most caption information out of a image like the core elements of art and photography.

I wish Sasori never found out about 4chan or this general

>>110280183
One good thing that could come out of constant trolling is that Sasori would fuck off with his autism, but yeah no such luck it seems.

>>110280183
sasuga
sashimi

dogshit general.

>>110280357
why?

>>110280364
Group of discord troons performed a hostile takeover and just use this thread for their 10-man circlejerk.

>>110280364
why wouldnt it be? literaly zero input and zero good sexy pics of women

>>110280183
who?

>>110280446
Go to a board that is designed for what you are looking for.

>>110280137
Stop fooling yourself:

-Krea 2 is a bad model, it can't produce good gens

>Your loras are equally bad and no one will pay for them

>The model is obsolete and no one uses it anymore

File: ElonVSEpstein.png (1.02 MB, 2048x2048)

1.02 MB PNG

>>110280379
Fuck this took up a lot of time, not doing that again.

File: 00000-1308421286.jpg (665 KB, 2880x1920)

665 KB JPG

>>110280497
honestly what are you trying to achieve was this pointless lying and trolling? I don't give a shit about what the normie masses or general mass of local ai enthusiast think of krea2. I was never into this hobby to make money and grift off people. I put the loaded time and effort to make loras that fit my wishes, needs and desires. Not to please random people on the internet. I come here to share my loras and gens to potential like minded anons who may like value my contribution. Not interested schizo drama and lowcow shit.
It's been nearly 5 months since the krea2 model weights got released and i haven't seen any other model come close to getting the results that i get with krea2. The model is far from being obsolete and dead. It's just that this model is not vramlet friendly (8-16gb vram) when it comes to lora training. You can't bullshit and rush the lora crafting and training process and hope to get expect decent results. It's a huge turn off to many of those prominent high status civitai lora bakers that play the SOE game for buzz.

>>110280719
kh1 kairi is my favourite <3

File: 6a099d4c9eb74ea7a7e1e369b(...).jpg (134 KB, 1920x1080)

134 KB JPG

>>110280719
Your kairi isnt even remotely correct. You've warped your perception of what a good gen by wasting your time training dead on arrival loras.

I dont care if a mode is "vram friendly for training" or not you wanna e-celeb grifter, I want a model like NAI that doesnt need me to tinker tranny my way with loras that dont even work half as well.

>>110280137
The default AI toolkit settings are bad and result on overbaked loras

>see nice h3 gen on civitai
>bro is using like 6x porn loras on top of the shitty eros max checkpoint
sigh

File: moss.jpg (699 KB, 1536x2048)

699 KB JPG

hey moss is pretty cool desu

you can just post existing artworks edited in klein 9b on civitai and they don't mind and everyone loves it. meanwhile i can't post my kino made in wangp coz no metadata or something

>>110279675
NTA but I've rawdogged Gemma 4 captions before and they have given me satisfactory results. I feel like for the most part unless I really need to get something right, the time sink of fixing mistakes probably isn't worth it if the captions are already like 99% correct. It compounds with the more images you have in a dataset too

File: ComfyUI_MOSS_02893_.png (3.35 MB, 1200x1600)

3.35 MB PNG

holy fucking KINO

>>110279691
sounds like it. im on 16g and its under a second per step. check your number of steps

>>110281094
Local models are kinda like funko pops. You get the base model and the collect a million different loras. Of course, on an individual level, these are just inane trinkets that you wont use for more than 10 mins.

Are there any leaderboards for image to text models (not to be confused with text to image)? The only ones i can find are for frontier cloud models or are very outdated/broken
I just want upload an image and have it described well enough for my LLM to understand it somewhat, particularly for japanese text. The "local" option for sillytavern comes with gpt-2 and its kinda dogshit

File: LLM AI visual reasoning M(...).png (94 KB, 1107x520)

94 KB PNG

>>110281485
There is no "caption bench".
Artificial Analysis MMMU and VIALS-AA are "visual intelligence" benchmarks in general which might correlate with caption quality.
I used Argon to caption a dataset recently and still had to make some edits. I dunno, especially if the dataset is small enough to sit through and make edits, it might be better to just caption with Gemma/Qwen MoE on your computer instead.
ChatGPT straight up safety blocks NSFW images right now, it used to just shy away from detail. Didn't even bother with Claude on that regard.

>>110280137
>nochekaiser
>not blacklisted since ponyxl
ngmi

Does an ez scripts fork for moss exist yet? I want to try baking one, although I only have a measly 16gb vram. I never trust one set and forget ones as it ends up shit to me (prodigy, the 1 LR ones, etc.) but I'm also a techlet and it's what I use for anima.

>>110281551
its a shame. i used to like them because they made good concepts for XL, but then they started making every lora cost 5k fucking buzz, and the quality took a nose dive. its like they dont test their loras at all.

>>110281609
>Does an ez scripts fork for moss exist yet?
sd-scripts backend supports it. There is probably an up to date fork for the GUI wrapper somewhere.
>I want to try baking one, although I only have a measly 16gb vram.
I am baking on 12gb, you would be fine.

Four consecutive replies to bump his thread... is still totally ignored. Ayy.

>>110281533
Its a little concerning how little downloads Gemma image to text models have on huggingface, Qwen seems a hundredfold more popular so ill try that first. Its a shame about the benchmarks because they seem to be doing image-text to text tests which probably wont work for image to text models. Someone should probably make one someday

how the hell do you autotag for moss loras

File: 17384858483.jpg (168 KB, 719x662)

168 KB JPG

Hello, I'm from the anime general.

The user Catjak seems to be involved in an artificial NAI vs Local war since weeks in our general, which never used to be a problem since we take no side between Local and SaaS.

>>110281404 name drops Catjak as the culprit.
>>110281475 looks like Catjak samefagging and gaslighting.
>>110281572 is from the same anon who got IP banned for name dropping Catjak as the culprit.

Pic related is a screencap of the range banned anon.

I don't want to accuse anybody, since this still isn't enough proof, but something is going on here.

>>110281731
This is catjak false flagging since he feels embarassed about his gens around NAI chads, we told him to report catjak to administators already we dont care.

>>110281648
Problem is finding out where that gui wrapper is and also which model, vae and TE to use when making a lora, all these convrot, bf16/8 and sft is really confusing where other models are just three singular things, if not i'll just keep waiting for the actual anime base one.

has anyone used bbox's with sdcpp? nothing is working for me

>>110281731
told you long ago that this nigger is in bed with the mod team. don't know if xhe's an actual mod though, the response times seem a bit slow, but xhe sure does run to his butt buddy to complain about stuff xhe doesn't like

this is too much fun. i was able to transform a cartoon character into a real life version with prompting alone using h3

File: chibicap.mp4 (3.39 MB, 1172x728)

3.39 MB MP4

>>110278953
Are you just training with the booru tags or are you feeding them to an LLM for json/natural language captions?
Here's chibicap assisting with reference image selection for face grounding.

>>110281961
NL captions with Gemma4. Booru tags are too noisy for Moss imo.
>face grounding smart move. saves so much time on training data prep.
are you using the raw json output or feeding it directly into sd-scripts?

>>110281874
Just learn to use the cli or ask claude to vibeslop you a GUI I dunno I do the former at least.
>which model
Base
>vae
flux 2 small
>TE
The one you use for genning, optionally download int8 convrot one but you can just quantize during training time.
>all these convrot, bf16/8
use --convrot_int8_text_encoder because you will get OOM on bf16 with 16gb. --cache_text_encoder_outputs_to_disk while we are at it.
Run the unet on bf16
>sft
Refer to above, forget it and use base for training
>i'll just keep waiting for the actual anime base one.
Suit yourself.

moss doesn't work when i replace the base text encoder with moss_te_heretic_full_int8_convrot.safetensors
to uncensor it.
can i uncensor it via a lora instead?

>>110281133
Best image model yet, despite being relatively small

I wonder if BFL will ever release another open model, they are so far behind and so sloppy censored their models are DOA

This has to be the schizo troll at this point, sorry out of daily (You) quota.

Someone is eager for this thread to die... that's not very polite. I can see you :D

File: Screenshot 2026-11-09 at (...).png (977 KB, 1702x925)

977 KB PNG

>>110279675
I pre-tag my images with character names and track it in a mysql db so at captioning time it can be fed to the LLM. The LLM then knows which characters are definitely in the image and what distinguishes them, it really improves the accuracy. I also flatten my datasets so main characters don't overwhelm supporting characters.
In picrel, some characters are "unknown" because the captioner is told if it does not get a character tag and distinguishing characteristics, it is not allowed to name them. This greatly reduces made-up assumptions about who is in an image.

>>110282041
He's not really good at it. Where is this janny lore coming from when we see the schizo break rules daily?

>>110282005
moss isnt censored, and using uncensored text encoders wouldnt work anyway because the image models arent trained on those sex terms.

>>110282163
It's a troll, he ask this constantly when he's pushing for bump limit. You can even Google the answer now.

>>110281961
>Are you just training with the booru tags
no tags, straight into the LLM for natural language captioning. But you're making me think I could probably save the tags as additional metadata to help the LLM.

>>110282063
moeshit for the local tinker trainy swine

how perfect

Is there a way to add negatives to H3 with NAG or something?

>tinker trainy

>>110282279
yeah, prompt
day 1 we figured out "she doesn't have long hair" results in short hair
I don't know how it works, it just does

>>110282163
nevermind, it probably just doesn't understand lesbian sex. ty anyway

>>110282177
no, you paranoid schizo fuck.

File: images.jpg (27 KB, 463x431)

27 KB JPG

>>110282212
>But you're making me think I could probably save the tags as additional metadata to help the LLM.
Are you fuckers dumb? Did you really not think of doing natural language + tag captioning until some anon mentioned it? Yes, it will improve your lora quality (it will still be local dogshit thoughbeit).