Blessed thread of frenship

Thank you debo.

>tune of h3
Meh

>>110291585
TANKS 4 BAKE ANON

I hate how much trial an error goes into baking a lora.

y no gen coleg

>mfw Resource news

11/10/2026

>MiniMax-H3-Image
https://huggingface.co/MiniMaxAI/MiniMax-H3-Image
https://huggingface.co/Comfy-Org/MiniMax-H3-Image

>LTX-2.5 Stage 1: 2-Step Distillation LoRA
https://huggingface.co/lvladikov/LTX-2.5-Stage1-Distill-2step-LoRA

>Decoupled Mode Connectivity for Base-to-Novel Generalization in Vision-Language Models
https://iemprog.github.io/DMC

>Video-Coder: Reinforcing Agentic Coding for Video Generation
https://cywang735.github.io/Video-Coder

>ReCaVSR - A one-step streaming video super-resolution built on Wan2.2
https://github.com/PedroMarinhoDev/ComfyUI-ReCaVSR

>H3 Editor: Timeline video editor for MiniMax H3, inside ComfyUI
https://github.com/igorls/comfyui-h3-editor

>H3MD Director: Seamless long takes with Chained Context Windows
https://github.com/1113800942/comfy-h3md-director

>MOSS-Image Control: Control LoRAs Collection for MOSS-Image-SFT
https://huggingface.co/collections/thedeoxen/moss-image-control

>moss_image_base_dmd2
https://huggingface.co/data-archetype/moss_image_base_dmd2

11/09/2026

>Open-H3-Regenerate-2K-IC: Consistent High-Resolution Video Refinement via In-Context Learning
https://github.com/NVlabs/Sana/tree/sol-engine/models/Open-H3-Regenerate-2K/IC
https://huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K-IC

>T2I Models Rarely Invert Negation but Often Fail to Suppress Negated Concepts: Benchmarking Negation Suppression in Text-to-Image Generation
https://github.com/IAB-IITJ/T2I-NSI

>MESAM: Enhancing SAM2 for Few-Shot Segmentation through Semantic-Visual Memory Generation and Refinement
https://github.com/tianyu-zou/MESAM

>Fizgig v9.3.0: Training maps, faster compile warm-up, and fine-tunes with the optimizer you pick
https://github.com/shootthesound/Fizgig/releases/tag/v9.3.0

>LTX x Chroma Awards Season 2
https://submit.chromaawards.com/pack/ltx

>HexAnima: Anime image generation that runs entirely on your phone's NPU
https://github.com/ayanoby/hexanima

>mfw Research news

11/10/2026

>LTX-Video: Realtime Video Latent Diffusion
https://ltx.io/technical-blog/ltx-video-realtime-video-latent-diffusion

>The RenAIssance: Caption-Guided Diffusion Models for Historical-Art Visual Media
https://arxiv.org/abs/2611.10599

>Disentangled Stylized Image Generation via IP-Cross-Attention Rearrangement
https://arxiv.org/abs/2611.10717

>ESTRA: Enabling Extreme Attention Sparsity for Training-Free Video Diffusion Acceleration
https://arxiv.org/abs/2611.10448

>AR-Edit: Training-Free Autoregressive Video Editing without Inversion
https://arxiv.org/abs/2611.09795

>VLM as an Editing Pilot: VLM-Steered Velocity Control for Flow-Based Image Editing
https://arxiv.org/abs/2611.10066

>Constrained Look-ahead Guidance for Interference-Aware Flow Editing
https://arxiv.org/abs/2611.10751

>On the universal structure in deterministic generative diffusion
https://arxiv.org/abs/2611.10338

>Beyond Prompt Attacks: Least-Action Auditing of Concept Erasure
https://arxiv.org/abs/2611.10106

>Split and Bridge: Multimodal Generation via Diffusion Bridging
https://arxiv.org/abs/2611.10441

>Semantic Synthesis over Brute-Force Search: LLM-Guided Transformation Design for Efficient Learning
https://arxiv.org/abs/2611.10940

>Improving the Diffusability of Motion Tokenizer
https://githubhgh.github.io/dimot

>STABLE: A Continual Learning Optimizer with Adaptive Drift Control
https://arxiv.org/abs/2611.10655

>PoseRieFlow: Riemannian Flow Matching for Whole-Body Text-to-Pose Generation
https://arxiv.org/abs/2611.09692

>CompositionEvidence: Controlling Geometric Evidence for Cinematic Composition Reasoning with Frozen Vision-Language Models
https://arxiv.org/abs/2611.10559

>MGE: A Membership Guided Data Extraction Attack against Diffusion Models
https://arxiv.org/abs/2611.10678

>BEiTScore++: Reference-free Image Captioning Evaluation with BEiT-3 and Conformalized Quantile Regression
https://arxiv.org/abs/2611.09505

>>110291775
thank you debo ^w^ <3333 very interesting links as usual

File: ComfyUI_temp_qddld_00023_.png (1.23 MB, 1024x1024)

1.23 MB PNG

rank 32, alpha 16, 1.5e-4
lineart is much better on this one but it's affecting the subjects too strongly. this cook just can't give teto her outfit.

>>110291704
Engines should already be firing because the way its shown, it would just smash down onto the landing pad.

File: ComfyUI_temp_qddld_00025_.png (1.27 MB, 1024x1024)

1.27 MB PNG

ok so it can.

>>110291807
if this is moss, can you share the recipe to make the lora? I want to try it.

so can it do 1-2 MP reliably? sometimes it's fine and sometimes i get clay smear grotesque out of it

I'm not impressed with the text to image so far but I have to remember comfy defaults are always shit tier

>>110291835
>the recipe to make the lora?
What do you mean?

>>110291855
my images so far with the default template are terrible
i now lurk until someone posts a catbox of an image that looks good

>>110291857
IF this is moss, how did you make your lora?

Verdict on the new h3 image model?

File: 2026-11-10-115443_1117x81(...).png (17 KB, 1117x812)

17 KB PNG

>>110291888
kohya-ss/sd-scripts

>>110291923
the lack of gens kinda speaks for itself.

>>110291928
thanks i'll try it out later i'm the one still waiting for a gui wrapper for this, if i copy paste it (adjusting file directory) in the terminal while sd script is running, will it work?

File: debo_pc_km_00029_.jpg (1.17 MB, 1792x1076)

1.17 MB JPG

>>110291797
:)

>>110291943
>still waiting for a gui wrapper
Vibecode your own in ten minutes nigga you're like my boomer dad who spends most of his day on his phone or PC but still calls me to help him install a Chrome extension when he could just google how to do it.

is H3-Image useless for anime/stylized illustrations? it eats limbs and other body parts on pretty much every try

example wfs that you pull directly from comfyui....It points to the video model which makes better outputs than the image model they dropped

>>110292025
you can get the actual t2i workflow here https://docs.comfy.org/tutorials/image/minimax/minimax-h3-image

or more directly https://github.com/Comfy-Org/workflow_templates/blob/main/templates/image_minimax_h3_image_t2i.json

i think minimax image SUCKS

alright havent been here in months
whats the current local sota for t2i and t2v?

>>110291990
Your dad calls you because he enjoys talking to you and values your skills with computers. He told me this after I fucked him btw

File: 2026-11-10-121919_986x485(...).png (12 KB, 986x485)

12 KB PNG

>>110291943
You need to clone the sd-script repo and install the dependencies.
What I'd recommend doing after is just making a bash script with what I pasted, that way you can just run it without having to remember everything.

You can read:
https://github.com/kohya-ss/sd-scripts/blob/main/docs/moss_image_train_network.md
(or have your clanker do it)

The only other missing part is you need to setup the toml for your dataset.
the dataset is just a folder with your images + a .txt file with the caption for the image.

File: MossTurb_plus_comfy_highR(...).jpg (1.55 MB, 2368x3552)

1.55 MB JPG

>>110292039
Thanks
After testing this model I have to say
It fucking sucks especially compared to MOSS

File: 1769571661340128.png (9 KB, 686x47)

9 KB PNG

People will get filtered by H3 Image

>>110292189
Even with LLM it's range is heavily bland most of us don't care about IRL shit, you can even use the image model wf that uses the video model, get the same size and cohesion and end up with more range.

lora training is cool.
I get free heating and kinos later.

>>110292168
And yet moss sucks too...

>>110292243
No it doesn't

stop falling for this bait.

oh the template made it to cumfy after all, it was just buried under 20 other minimaxes including a text to image that used the video model.
>minimax-h3-image

anyway, it sucks and updating comfy to support it broke some of my custom nodes

>>110292256
It really truly does (for anime)

>>110292039
Hmm, the fields in the template are all mismatched for me. (The model filename is in the VAE field, the sampler is in the unet field, the scheduler is in the clip field, etc.) I guess I'll wait for them to unfuck things.

Tesla should have been below 100

>>110292288
jesus fuck, anime AI gen has become a solved problem now, the fuck more you want?

>>110292288
No shit? They didn't train on booru images at all. That's why the devs said they're going to train a separate anime model. Stop seeing nails everywhere just because you hold a hammer.

File: lol.jpg (45 KB, 449x401)

45 KB JPG

>he still prompts using booru tags
>in 2026

>>110292308
I never held a hammer, instead I've held in my hands the philosopher's stone that is NAI.

>itto justo needsu da booru prease understandu
You parrot stuff you hear about, but booru is not enough by itself to train a proper model, its just a tiny portion of whatever any artist has decides to make public via shitter or whatever else. Please understand our standards are way different.

>>110292348
you're on 4chan lol, quit acting like you have standards

>>110292388
As opposed to what, trooncord or civitai?

File: MossTurb_plus_comfy_highR(...).jpg (1.5 MB, 3344x2512)

1.5 MB JPG

So with a trial
New Kromoss has significantly better color
Needs a lora update because this family of models go to hell when not using a lora from the same model, might be the best candidate of lokr
I think the model is going in the right direction but holy fuck stop half releasing shit.

>>110292395
fuck off, shit I like VS shit you like is not debate worth having

>>110292423
ok but why should I use this if I'm not in dog dicks

File: file.png (456 KB, 864x517)

456 KB PNG

new toy

>>110292461
It makes better loras compared to base moss

>>110292485
>*citation needed

Where's the alt collage? OP's is shit

File: debo_pc_km_00030_.jpg (1.23 MB, 1792x1076)

1.23 MB JPG

>>110292189
people who use h3 mostly have used LLMs to write their prompts. they're pre-conditioned to do the same for H3-image

>>110292643
I didn't ask you

>minimax h3 image
>huge model, even huger text encoder, resource intensive to train
>except you can't train it at all, because they only released a 6-step distilled turbo version
>doesn't really have NSFW knowledge like MOSS
>in terms of raw prompt comprehension, still worse than krea
>VAE is complete and utter ass, like SDXL level, they cope by requiring the model to gen at massive resolutions
>same shit license as the video model, you can't even display the images if you're in the US or EU
I know it's a meme phrase but honestly the model is DoA. like what does this do that Krea / MOSS don't. and nobody is gonna train it

File: Ether Jungle.jpg (725 KB, 1280x1152)

725 KB JPG

>>110292643
All those images have a pretty nice color palette and composition, they really look a videogame map. Until you look closely and notice the huge amount of little borked details. They would be super notorious in a game where you have too look at the same image for a good while. Still, pretty impressive, local models had advanced nicely. Maybe a map where the camera is not so far away will work better. Something like pic related -not a gen-.

>>110292577
My training, you can try it or not I don't care
>>110292758
it's fucking garbage

File: debo_pc_km_00032_.jpg (1.29 MB, 1792x1076)

1.29 MB JPG

>>110292758
>>except you can't train it at all, because they only released a 6-step distilled turbo version
people have been training on distilled models for a while now. this isn't a blocker

>>110292765
yea, def a lot of minor mistakes and artifacts in these. I could surely make these cleaners if I mess around with post-processing nodes but I liked these plenty as-is. they make me imagine what these games would actually play/feel like

>>110292813
You are out of your depth here, nobody is going to do that for a mediocre bloated model.

>>110292813
Yeah we know you can't produce non slop

>>110292893
> we
stfu cumjack

>>110292920
Can you stop?

File: 9eb802eb-53eb-444e-977f-c(...).png (2.8 MB, 1536x1024)

2.8 MB PNG

>>110292813
This is a interpretation of an old chat GPT image model of a new map of the same game, good details but the texture is lost.


>>110292893
What do you have against him? Where did he touch you?

Also, be less predictable and respond with something different to accusing me of being him.

File: debo_pc_km_00037_.png (3.91 MB, 1792x1076)

3.91 MB PNG

>>110292840
nobody? there seems like there's always someone out there willing to do stuff like like

>>110292952
still a very cozy gen. a banished-style builder in feudal japan would be really cool

>>110293000
That was during the blank era, that model performs worse than models a fraction of it's size, get real.

>>110293000
That is a very bizarre use of the word cozy

>nigbo is lonely again

File: debo_pc_km_00039_.jpg (1.3 MB, 1792x1076)

1.3 MB JPG

>>110293026
aesthetic city builders are pretty cozy

After trying out uncensored MOSS pussies I can't go back to mangled pussies like H3 Image.

>>110293054
>>110292920

>>110293065
this is pure slop, there's no memorizable building

>>110293079
he does that on purpose, ignore him.

>>110293075
Not your boogeyman
I just don't like nigbo spamming the second general into the ground

>>110293116

He is posting AI images in a thread about AI generated stuff... meanwhile you are just bitching.

>>110293085
We are so back

>>110292758
B-but its so big, anon! It can't be bad if it's so big!

>>110293116
He does that daily already
Filter him, he'll anon seethe anyways

>>110293085
did you recaption for moss or? idr what you usually use to caption

>>110292637
I'm sure next time you'll get featured, anon. I'm sure of it.

File: minimax_h3_image_t2i_00002_.png (3.51 MB, 1536x2720)

3.51 MB PNG

4mp image in 32 seconds, not bad.

File: moss_image_sft_int8_convr(...).jpg (319 KB, 1024x1504)

319 KB JPG

>>110293201
I dumped images to civitai trainer and let it do the captions.

File: ComfyUI_00589_.png (1.61 MB, 1088x1928)

1.61 MB PNG

>>110293227
Same prompt with Krea2.

>>110293231
oh i thought you were the anon who baked s oyjak loras for previous models

>>110293227
>>110293238
Can you make them look non generated?

File: 2026-11-10-141354_965x350(...).png (245 KB, 965x350)

245 KB PNG

>>110293238

>>110293277
welcome back SDXL

File: 1788191620277933.jpg (2.97 MB, 5472x3072)

2.97 MB JPG

wow this model fucking sucks

>>110293116
> Not your boogeyman
yeah
sure

>>110293277
for >>110293227
wrong image.

I put my full name, address, and social security number into moss and it generated a 1:1 image of me

>>110292481
Digging it so far. It has the variation and crusty sharpness that I like. Bloated as fuck and seems to lock up my computer more than the fucking video model, kek, but lowering the res works fine. Going to do more testing then move on to editing tomorrow.

>>110293300
The fact that these are so sloppy adds to the rage inducing aspects of these images.

File: 6SFl8FdAOLs.jpg (56 KB, 376x603)

56 KB JPG

Is it illegal to post a video of bill gates getting fpv droned?

You didn't gen it so yes it is illegal to post.

>>110291683
And its nondeterministic so you could actually be making it worse without realizing

File: Krea2_turbo_00401_.png (2.04 MB, 1088x1928)

2.04 MB PNG

I'm not getting good results with complex prompts meant for Krea2, moderately complex prompts look decent. Using the recommended 16mp doesn't help at all. Simple frontal nudity works Ok.


>>110293254
Probably, but my autism centers around correct anatomy and not borked details. I don't see anything wrong with Krea2 version >>110293238 If your refer to the lack of expression and neutral pose, I prompted exactly for that. Also the prompt was optimized for Krea2.

>>110293359
is the 6pack prompted for?

File: minimax_h3_image_t2i_00069_.png (3.32 MB, 1408x1856)

3.32 MB PNG

the prompt enhancer is replacing retarded with [redacted]

Does inpaint still lurk here? I have question about The hh48288196 lora.
What made you target only layers 20-34 ?

>>110293359
can we get your prompt pls boss she's a cutie
not surprised h3 img stinks, many such cases with 99% of these chinese releases.
>he says with MOSS hidden behind his back

>>110293395
lmao it really is DOA

>>110293419
>can we get your prompt pls boss she's a cutie
You're unable to derive the clearly simple prompt from anons gens? Baka newfag

>>110293438
>He doesn't understand how special individual prompt builder's prompts really are
maybe, have you considered, riddle me this, that you may be the "Baka Newfag".

>>110293449
>special
>its just a girl with a black bob and bangs
kek

>>110293419
You are a fucking moron

>>110293340
>>110293379
uhhhh she looks minor-coded bro

>>110293459
You just don't get it..
I bet those homos in maid dresses are more your thing.

>>110293475
Confucius say, you are the moron.

File: minimax_h3_image_t2i_00022_.png (2.5 MB, 1088x1920)

2.5 MB PNG

>>110293384
Yes, but it's mean to be subtle, like in the Krea2 version. MiniMax image make her super jacked. But to be fair to MiniMax, Krea2 was not giving her any abs so I had to strongly empathize that aspect in my prompt. Things are not looking good for MiniMax image. Higher resolutions doesn't give me a better and more detailed image, just bigger blurriness and more noise

>>110293476
Fuck off, zoom zoom faggot.

>>110293069
Who has the balls anymore to train on completely uncensored pics from the internet? Moss made this the default expected behavior of any new open model des.

>>110293476
it's best not to project and give away your inner thoughts

>>110293414
those are the good layers that's why

>>110293510
Isn't there an argument to be made that as long as no human views the images it's technically not illegal?

>>110293532
>>110293501
How can it be so big yet so bad?

Could someone point me to the last, to date, workflow using convrot-pruned-int8 for Minimax H3 after the "Dual clock Eluer" change?
Is "Spectrum Apply" still a thing?

utterly raped by MOSS

>>110293501
>>110293532
that is fucking rough

>>110293528
The only thing that's illegal right now in freedomland is disseminating deepfakes, possession is fine (some will claim possession is illegal but everyone whose been arrested and charged was because they also had real CP). It's kinda like how weed is in something states.
>as long as no human views the images it's technically not illegal?
Interesting to bring up because one of the many anti ai arguments is that it doesn't matter if no image is generated, the mere act of training (not just specific loras, but training anyone at all into any model at any stage) without "consent" should be charged as sexual assault. Thankfully that hasn't been enacted into law quite yet.

>>110293546
dumbass labs are still catching up to china, they have yet to understand that all you need is real data no synthslop

File: ComfyUI_00723_.png (2.23 MB, 1088x1928)

2.23 MB PNG

Complex Krea2 prompts don't work at all in MiniMax image. I may try to convert my prompt later... but they are 500 words long...

>>110293315
Only if you live in a first world dystopia like the United Kalifate of Bongistan

>>110293532
Good fucking question, seems to be pretty bad.

File: p0ml3h.jpg (805 KB, 2176x2688)

805 KB JPG

>>110293562
even anima rapes it imo.

Why haven't you switched to MOSS yet, anon? What's stopping you?

>>110293501
These thin skinned, limp wristed, broccoli hair permed fucks would call their moms if they had to witness the CP spam 4chan used to be bombed with not so long ago.
I miss that, these weak retards wouldn't be here to begin with.

what do you do in krea and moss once you're finished spamming the generic realistic 1girls?

>>110293599
>Only if you live in a first world dystopia like the United Kalifate of Bongistan
It's also illegal in the US.

>>110293645
train lora.

>>110293476
>minor-coded
sociopathic tumblrisms used everywhere make me sad

>>110293600
pussy

>tfw you see your gen reposted in another regular nonai thread
>tfw you see multiple of your gens posted to various nonai threads or used as the OP pic for various nonai threads for over three years
never gets old

File: lora.jpg (3.02 MB, 3552x2368)

3.02 MB JPG

>>110293395
Absolute dogshit
When you censor words your model deserves to get pissed on

>>110293680
example?

>>110293510
>Moss made this the default expected behavior of any new open model des.
AFAIK it's been the default in most closed labs since quite a while, they just censor you by analyzing the prompt and resulting image.
Which is actually less stupid than completely destroying any understanding of this stuff, which can be anywhere between nudity by itself (the most retarded by far) to genitals to nsfw/intercourse.

>>110293645
start spamming landscape or abstract kino

>>110293680
what's wrong with it?

>>110293419
>>110293683
Do we have newfags here seriously factoring in the prompt rewriter when judging a model? And they aren't immediately getting laughed out of the thread?

>>110293683
but can it write cum?

>>110293722
Nothing is wrong with it, I enjoy when my sovlkino gets reposted. I'll see them used as wallpapers in the desktop threads as well.

>>110293395
lmao, use an anti refusal lora for the text encoder and it should be fine

>>110293740
ive slowly come to the realization that its better to let them fuck up as even if they are corrected they shut their ears and ignore

>>110293740
I used the prompt rewiter and multiple llms. What the fuck are you even saying?

File: minimax_h3_image_t2i_00032_.png (3.12 MB, 1088x1920)

3.12 MB PNG

>>110293599
Meanwhile MiniMax image... The 4mp images didn't get borked hands, but rest was as bad as this one.

>Maximum file size is 4 MB
Stupid 4chen...

>>110293644
And if only they were so fucking stupid, they wouldn't be so oblivious to the direct chain of causality between their declining life quality and the promotion of weakness (of will) and stupidity.

Yes, your love for censorship shares the same root as to why you will never own a home.

>>110293753
I feel like there's always a 50/50 split between someone just trolling and being genuinely retarded.

>>110293745
why would i when another model can do it without some cope lora?

I lost faith in the Minimax team after Minimax Music 3. No finetuning it, and they weren't willing to answer questions, plus it was more slopped than H3's short snippets so they held back. Shows they weren't truly about the open spirit nor willing to contribute, H3 was their best contribution but then it's only a publicity stunt from then on.

did the minimax team get forced into rushing out slop because of moss or why did they release such a dud?

>>110293740
the thread was a shitshow for too long for any newfag to learn anything

MOSS-video when?

>>110293762
The Great Erasure Of Nipples

>>110293629
Not a pedo.

>>Maximum file size is 4 MB
>not using xt
newfaaaaaaag

>>110293800
looks like zoomers are awake, having fun playing pedo hunters?

>>110292146
t2i is moss or kromoss
t2v unironically ltx but i2v/ref2v is h3 hands down

>>110293793
We'd all die.

>>110293787
>rushing out
that shit's been worked on for months fuck you mean? its just another bad release in a sea of bad releases.

>>110293793
I'd rather them stay in their image making lane and deliver a better uncensored image model, or their anime finetune, than scatter and use resources on new stuff.

File: twoworlds_.jpg (3.55 MB, 3552x2368)

3.55 MB JPG

>>110293740
I have no idea what you're talking about but I promise you my prompts will make you sweat

>>110293769
Yeah, it was clear from how much publicity H3 got that they'd never release anything good ever again

>>110293827
cheers

>>110293740
>Do we have newfags here seriously factoring in the prompt rewriter when judging a model?
You would not believe the number of fags blaming models because they suck at prompting or are unable to read the notes from the author. It's an epidemic.

File: Kromoss_Singerlora_00017_.jpg (3.83 MB, 3552x2368)

3.83 MB JPG

Needed the texture bump on this image

>>110293845
It's incredible how well the heuristic works for image/video models :
- release good model
- people talk about you
- promise new cool stuff (remember h3 promising to release their upscaler and sparse attention configuration?)
- never fulfill any promise, either disappear or release disappointing things instead

is there any kromoss version that isn't turbo (8 steps) and cfg forced to 1?

>>110293896
They all work best with it and negpip from the kj node actually makes it great

krea and minimax teams have committed mass suicide BREAKING NEWS

File: minimax_h3_image_t2i_00039_.png (2.05 MB, 1088x1920)

2.05 MB PNG

>>110293869
Enhancer or not, I get pure borked gens. However, I admit that I have only tested my Krea2 prompts, which are long and pretty detailed. Simple prompts work better, they are still bad, just not as a bad.

how do you make moss actually output good quality images though

very carefully

>>110293956
{
"width": 3552,
"height": 2368,
"bbox_coord_space": "abs",
"bbox_format": "xyxy",
"summary": {
"meta": ["@singer"],
"canvas": "single",
"caption": "Oil painting of two Kirby characters standing side-by-side in a sunlit room overflowing with colorful flowers, contrasting an innocent Kirby with an angry, armed cowboy Kirby."
},
"items": [
{
"label": "character",
"caption": "A classic pink Kirby with a soft, matte texture, featuring a wide, innocent smile and large, sparkling black eyes, looking at the other kirby with a curious expression.",
"bbox": [1024, 960, 1792, 1856]
},
{
"label": "character",
"caption": "A second pink Kirby standing in a rigid, aggressive posture with a fierce scowl, deeply furrowed brows, and a downturned mouth.",
"bbox": [1792, 960, 2560, 1856]
},
{
"label": "object",
"caption": "A weathered brown leather cowboy hat with a pinched crown and wide brim, perched on the head of the angry Kirby.",
"bbox": [1856, 768, 2432, 960]
},
{
"label": "object",
"caption": "A small, matte-black semi-automatic pistol held firmly in the right hand of the angry Kirby.",
"bbox": [2432, 1216, 2720, 1408]
},
{
"label": "object",
"caption": "A wooden flagpole topped with a fabric American flag, held in the left hand of the angry Kirby, with the stars and stripes draping slightly.",
"bbox": [1728, 832, 2048, 1792]
},
{
"label": "object",
"caption": "A dense carpet of vibrant red tulips, yellow daisies, and white lilies covering the entire floor and clustering against the walls.",
"bbox": [0, 1536, 3552, 2368]
},
{
"label": "object",
"caption": "Off-white interior walls of a room, softly lit and partially obscured by tall flower stems at the edges.",
"bbox": [0, 0, 3552, 1536]
}
]
}
(low quality: -1.0), (worst quality: -1.0), (blurry: -1.0),

Comfy refuses to load the RTXSuperResolution nod after an upgrade.
Halp.
It's the only one I didn't manage to solve after my workflow got literally raped by the changes in comfy.

I appreciate the anon posting a bunch of 1girl, blank expression, standing gens. Very important to know if a model can do this very complex prompt.

that one meme of all the dudes shotgunning their brains out then the guy looking on at them like they're stupid, moss's team is the guy not shotgunning his brains out
but knock on wood, their next release will be dogshit like everyone else's after they take a whiff of success.

im wondering how you get moss to make shit images desu. some anon the other day was posting moss gens that looked so (ai-generated:1.2) it was hard to believe

>>110294005
Use the wf comfy gave to everyone

>>110293982
This, but unironically.

>>110293982
>Very important to know if a model can do this very complex prompt.
Yeah, too bad it can't but hopefully that means it can still do simpler prompts.

>>110293998
Don't worry, every time they tease something the retards will post about it here and bash you for not believing everything they say without a second thought

>>110294008
thankfully im not that retarded

>>110293982
I agree, no irony.

>>110293977
>using the meta value
Based

wake md up when there's something that can generate these 1girls in realtime according to whatever we're talking about

yawn

wake me up before you go go

So minimax image generator dead on arrival? Its hard to tell from the reddit bots and shills

File: image_minimax_h3_image_ed(...).png (1.68 MB, 1024x1024)

1.68 MB PNG

oof

>>110294078
At least give it a chance by using the correct resolution range.
>The effective working range of MiniMax-H3-Image is about 4 MP to 20 MP, and it generally gives the best results at around 16 MP. Common 16 MP sizes are listed below:

>>110294072
Why do you care if you cant run it yourself?

presenting my vibe-code pice of ai-shit
https://github.com/Koreprobro/nerv-studio
a bridge between artificial idiot and diffusion models.
XOR VRAM logic. best for multi-gpu owners.

https://www.youtube.com/@adrianimmorales - several debugging short video generations. quality is heavy dirty, draft quick generation way.

ALL made LOCAL.

KI-7 - about 4 hours of fully autonomous work of qwen3.8 + diffusion generation. without detailed instructions. result 10min video

>>110294092
I mean by that highly autistic corpospeak are we positive that means it's MINIMUM 4MP? I find that hard to believe.
But then again i'm finding it relatively easy to believe that this thread's examples demonstrate its dogshittiness so i guess i should habeeb it.

>>110294105
Interesting. Will look into this.

>>110294105
harvesting your ips

>>110293977
Imagine actually prompting like this. Might as well pick up a pencil.

File: 1784677776056456.jpg (114 KB, 1713x970)

114 KB JPG

>>110294117
At least worth trying, I'll try it myself this weekend, I hope that at least the edit capabilities are worth it.

>>110294078
AAAAAAIIIIIIIIIIIIIIIIIIEEEEEEEEEEEE

>>110294135
>he thinks anon writes all that by hand
Newfren pls

>>110294072
the best advice is to do the opposite of whatever plebbit recommends

File: MossTurb_plus_comfy_highR(...).jpg (2.12 MB, 2368x3552)

2.12 MB JPG

>>110294135
Are you serious anon?
>>110294157
I like to think they don't have enough vram to run a llm that can help them.

>try out 16 mbp
>takes 5min for the image
>still uncanny 3D render dogshit
>too big to even post on 4channel

>>110294135
Do you think writing six sentences is a high effort task?

>>110294184
>>110293801

>>110294140
>2026 model
>those resolution restrictions
PFFFFTTTTT LMAO

>>110294186
finally, dall-e 3 at home

I am NOT drawing rectangles for your shitty built-in regional prompter.

>>110294092
I'm doing that, it helps, but not much.

>Professional photograph
>Shot on a full-frame camera with an 85mm lens at f/2.8.
>Studio lighting.
>Realistic.
>Photorealistic, real photograph, not an illustration.

MiniMax Image: I'm going to gen a 2.5 image and you are gonna like it.

>>>/hr/5304898

I'm going to try to prompt from zero, Krea2 prompts were a failure.

>>110294243
geg

>>110294243
it knows nipples??

File: mastery.jpg (3.2 MB, 2512x3344)

3.2 MB JPG

>>110294235
I would never use that clunky shit

I am using ai for drawing rectangles for your useful built-in regional prompter.

so is minimax image any good

>>110294420
good at being bad

>>110294432
kek

File: paint_.jpg (3.04 MB, 3344x2512)

3.04 MB JPG

>>110294420
It's absolute turbo distilled dog shit

File: image_minimax_h3_image_ed(...).png (1.68 MB, 1376x768)

1.68 MB PNG

so far h3 has been the only edit model able to do this shitty edit properly. didnt try fibo but qwen and flux3 klein wasnt able to add alice in

File: t2i_00001_.jpg (1.69 MB, 3360x5024)

1.69 MB JPG

>>110294420
The only thing that looked good so far is the default prompt. People look fried as fuck, even the video model is better at genning images of people than this. I'll give it that it's fast as fuck and you can gen huge images but that's it. Will have to experiment more.

anyone got a good moss flow they could share?
I look at the ones on civit and they have a real fixation on adding a million nodes that really don't do anything

>>110294469
>gen images with moss or krea 2
>edit images with minimax
like this I guess, though the lack of nsfw understanding probably means even editing will reach a limit

>>110294492
What's the intention behind this when models exist that can give each element purpose?
Why are you unable to use any of the most basic tools?

>>110294502
I use the default comfy provided one, it works ok.

File: image_minimax_h3_image_ed(...).jpg (1.28 MB, 1664x2496)

1.28 MB JPG

editing with 4MP images is slow

File: paint3.jpg (3.43 MB, 3344x2512)

3.43 MB JPG

>>110294502
Skip layer 22, 24, and 26

File: debo_pc_km_00042_.png (3.94 MB, 1792x1076)

3.94 MB PNG

>>110294504
child, stop crying

>>110294072
how about you try the model out for your self and form your own verdict. I swear this place is just reddit 2, attacking anyone who isnt exclusively using or who dislikes krea2 or moss.

>>110294522
why

>>110294535
unlucky numbers in chinese culture

>>110294502
Just use the default one with whichever lora loader you fancy

>>110294528
I'm still amazed by your inability to do basic things
>>110294535
I got better results from doing it the model says skipping those layers do improve things but there is a chance of issues but I never encountered it with the dmd turbo

>>110294510
>>110294546
aight, I'll trust the comfy default flow, this one time!

thanks lads

>>110294549
ah ok

>>110294528
every time i zoom in to this style of photo, the small details are always incomprehensable trash

>>110294522
dont skip 24 and 26, in fact dont skip anything at all if you want something that looks off and ugly (on purpose). treat skiplayer as if you are adding .5 to your cfg

File: reprise_.jpg (2.2 MB, 2896x2176)

2.2 MB JPG

>>110294606
works for me, care to show examples

File: debo_pc_km_00043_.png (3.9 MB, 1792x1076)

3.9 MB PNG

>>110294602
wow you're really out here struggling

why is he arguing with himself as an excuse to post more pictures

File: Screenshot 2026-11-10 163244.jpg (714 KB, 3792x1966)

714 KB JPG

>>110294420
it needs a second pass hi res fix support. i hope it gets forge neo support. you can try it out on wan2gp. results are not crisp and ultra sharp like krea2.

>procure a nice data set with varying resolutions (512 - 1536)
>civit HARDCAPS the resolution to 1024
WHAT THE FUCK IS THIS BUZZ GOOD FOR
HOLLYYYYYYYYY HSITT TY
REEEEEEEEEEEEEEE

>>110294712
sell your buzz and buy a GPU

>>110294690
comfyui default workflow upscales the reference image to 4MP before feeding it to the model

File: MossTurb_plus_comfy_highR(...).jpg (1.92 MB, 2896x2176)

1.92 MB JPG

>>110294651
I think you are, you have put in more time than any of us combined and it looks worse than your sd 1.5 days.
Do you need a link to your model's documentation
You lost or something?

why are text encoders so big now?

>>110294760
Compared to what, CLIP?

File: 2026-11-10-16h47m11s_seed(...).jpg (1.41 MB, 2176x3840)

1.41 MB JPG

>>110294746
is there any benefit going above 6 steps?
i can't change the step count on wan2gp.
hate how restrictive wan2gp can be with certain image model unless deepbeepmeep deems it worthy adding features in.

>>110294817
No it still looks like shit

Why are MiniMax image, images, so crispy-crunchy!?

>>110294867
not enough bruises

>>110294690
Btw quick question does the government give you money for your autism or is it your parents or what's the situation like

File: ComfyUI_temp_airrp_00004_.png (2.12 MB, 1600x1088)

2.12 MB PNG

How do I unfuck the thing moss does when it looks like the pic is obviously stitched from regions?

File: MiniMax-Image_0016.jpg (2.53 MB, 2048x2048)

2.53 MB JPG

>>110294817
I tried 8 to 10 steps and not much benefit really.

>>110294835
It is pretty crispy, huh. Try the lcm sampler.

>>110294712
shoo poorfag shoo

>>110294817
>i can't change the step count on wan2gp.
Lol what's the point of that front end if it lacks basic features? That's crazy

>>>110294949
What is moss?

>>110294993
Moss is the Boss

https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT

File: 1782682906505475.png (2.59 MB, 1248x1664)

2.59 MB PNG

>>110294874
You are right, I should also see how good Moss is for ryona.

>>110295001
mogged by kromoss

>>110295001
>pointing a newfag to SFT
anon...

File: trip-frog2.png (1.87 MB, 1024x1024)

1.87 MB PNG

>>110295001
Thanks. It looks a bit overkill for just meming. But I will try to get it running. It sounds like it will probably be slow on a 3060

File: 2026-11-10-17h12m05s_seed(...).jpg (1.98 MB, 2176x3840)

1.98 MB JPG

>>110294928
the model is soo janky. not even as sharp krea2 and the seed diversity value is a lot higher but not as unstable as MOSS.

>>110295034
>It looks a bit overkill for just meming
desu you can make hyper specific memes and specifications with bbox prompting, also gore and coom

>>110295054
Aight, imma get this thing going. Thank you.

people just using the base MOSS or is there something else actually worth using?

dis nigga still yapping about krea kek

>>110295083
some use kromoss but desu moss base has more than enough in it already IMO

File: ComfyUI_temp_airrp_00014_.jpg (394 KB, 4096x1712)

394 KB JPG

I can just pick any number for resolution.

>>110295083
>>110295019

File: debo_pc_km_00046_.png (3.94 MB, 1792x1076)

3.94 MB PNG

what do we call minimax h3 image? h3i?

>>110295213
shit

you should ask that question in your dedicated thread instead of here tardbo

>>110295279
>:3
I don't know who that is, I don't read these threads, it's immediately obvious that you're desperate for drama.

>>110295287
Color tattoos on a black and white body, nice. I'm going to steal that concept.

>>110295279
You are the only trying to start a stupid, meaningless and boring internet fight.

>>110295325
try out the spot color tag

>>110295083
I'm also baking my own loras on moss base.

>110295279
Go ahead and bake a couple of troll threads that no one uses and eventually gets deleted by mods