based thread kill ani

>mfw Resource news

11/11/2026

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

>LTX-2.5 Stage 1: 2-Step Distillation LoRA
https://huggingface.co/lvladikov/LTX-2.5-Stage2-Distill-5step-LoRAs

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

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

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

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

>H3MD Director: Seamless long takes with Chained Context Windows
https://github.com/1113800942/comfy-h3-directors

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

>moss_image_base_dmd2
https://huggingface.co/data-archetype/moss_image_dmd2

11/10/2026

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

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

>Fizgig v9.3.0: Training maps, faster compile warm-up, and fine-tunes with the optimizer you pick
https://github.com/shootthesound/Fizgig/release/tags/v9.4.0

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

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

>mfw Research news

11/11/2026

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

>The RenAIssance: Caption-Guided Diffusion Models for Historical-Art Visual Media
https://arxiv.org/abs/2611.105

>Disentangled Stylized Image Generation via IP-Cross-Attention Rearrangement
https://arxiv.org/abs/2611.107

>ESTRA: Enabling Extreme Attention Sparsity for Training-Free Video Diffusion Acceleration
https://arxiv.org/abs/2611.104

>AR-Edit: Training-Free Autoregressive Video Editing without Inversion
https://arxiv.org/abs/2611.097

>VLM as an Editing Pilot: VLM-Steered Velocity Control for Flow-Based Image Editing
https://arxiv.org/abs/2611.100

>Constrained Look-ahead Guidance for Interference-Aware Flow Editing
https://arxiv.org/abs/2611.107

>On the universal structure in deterministic generative diffusion
https://arxiv.org/abs/2611.103

>Beyond Prompt Attacks: Least-Action Auditing of Concept Erasure
https://arxiv.org/abs/2611.101

>Split and Bridge: Multimodal Generation via Diffusion Bridging
https://arxiv.org/abs/2611.104

>Semantic Synthesis over Brute-Force Search: LLM-Guided Transformation Design for Efficient Learning
https://arxiv.org/abs/2611.109

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

>STABLE: A Continual Learning Optimizer with Adaptive Drift Control
https://arxiv.org/abs/2611.106

>PoseRieFlow: Riemannian Flow Matching for Whole-Body Text-to-Pose Generation
https://arxiv.org/abs/2611.096

>CompositionEvidence: Controlling Geometric Evidence for Cinematic Composition Reasoning with Frozen Vision-Language Models
https://arxiv.org/abs/2611.105

>MGE: A Membership Guided Data Extraction Attack against Diffusion Models
https://arxiv.org/abs/2611.106

>BEiTScore++: Reference-free Image Captioning Evaluation with BEiT-3 and Conformalized Quantile Regression
https://arxiv.org/abs/2611.095

What's the easiest way to handle the bbox stuff with moss? No one is doing them by hand right? RIGHT?

I wanted to say hi to all anons, I hope you have a good time here and a good day as well

>>110300723
llm for initial generation
and there are some tools on github as well

>>110300898
Can you explain that like i'm retarded? Because i am

File: 15726819_1072057402920504(...).jpg (90 KB, 960x582)

90 KB JPG

What's a good workflow setup for making a 100+ page comic book?
The characters and scenes need a consistent look. I dont want inconsistencies from chapter to chapter or from one book to the next.
Hopefully it can frame the content in the heroic Marvel style and not just 6 predictable panels every page from the same camera angle.
Some characters may be based on real people. The style will borrow from comic books of the past. Maybe 1970s Marvel. The characters can't be nerfed to look a little off, because the AI is avoiding intellectual property infringement. (picrel)
I'm a noob btw. All tips and tricks are appreciated. I will be updating my PC to help with this task, but I cant afford a $2K graphics card.

>>110300907
Well, I'm no genius either. I just ask some random free llm to follow the json structure on the moss-image site. It does sometimes. It needs to know what res you're genning at.
Seach for "moss-image bbox github" and you will get comfy plugins and what not i haven't tried them.

>>110300168
Is there any alternative to the noodle nightmare when i want to run the latest stuff? It's really uncomfortable

>>110301021
what pc do you have now? local models can run on pretty much anything so you should try them out before deciding to upgrade

>>110300568
>>110300568
>>110300568
real thread here

ty for bake

Blessed thread of frenship

>>110301177
>>110301092
let's go here, better fren bake
>>110300568

>>110301064
I am due for an update. We're looking for Black Friday deals.
I will add my question to the proper thread >>110300568

File: ComfyUI_temp_psfhf_00073_.png (1.06 MB, 832x1248)

1.06 MB PNG

Oekaki lora
https://civitai.red/models/3046271/moss-oekakidoodle-style
https://gofile.io/d/bq54M0gR

I'll probably revisit it at some point I think it's good enough for now.

>>110301215
Wrong thread, we're all at the fren bake >>110300568

>>110301215
thanks gonna test it soon, have you test it against kromoss?

>>110301275
have not. I only use SFT. It's trained on base tho.

>still samefagging his shitbake.

Danbooru is back.

>>110301836
I'm not gay but this makes me feel things.

fucking bounding box magic

>>110301707
>>110301836
>>110301851
>>110301891
we're all at the fren bake. join us bros
>>110300568

File: file.png (1.17 MB, 832x1248)

1.17 MB PNG

>>110301306
seems to work well enough

>>110301918
nice!

>>110300168
TANKS 4 BAKE ANON

>>110301707
thank fucking god

>>110301215
Based and ty anon

>>110300168
Thank you for baking this thread, anon
>>110301177
Thank you for blessing this thread, anon

Is there a model that does "Recreate image A in the style of image B" well yet?

>>110302097
*sniffs*

File: debo_news_00038_.png (2.62 MB, 1792x1312)

2.62 MB PNG

>>110300168
>mfw REAL Resource news

11/11/2026

>PAN 2: General-purpose world model by the Institute of Foundation Models
https://huggingface.co/IFM/PAN2

>pi-H3: Video Generation in 2–4 Steps Without a VAE Decoder
https://hanshengchen.com/blog/pi-h3

>Qwen3.8-27B GGUFs
https://unsloth.ai/docs/models/qwen3.8

>UniWeaver: Harnessing Unified Multimodal Models for Interleaved Text-Image Generation
https://poscape.github.io/UniWeaver

>AnyBreak: Towards Jailbreaking Any T2I Model for Any Concept
https://github.com/luqi-glory/AnyBreak

>What kills $v$-prediction? A Patch-wise PCA Perspective on Pixel-Space Flow Matching
https://github.com/THUMAI-Lab/RiT

>OpticalRAG: Pixel-Space Compression for Token-Efficient Retrieval-Augmented Generation
https://github.com/brighterr/OpticalRAG

>ContraSafe-VL: A Contrastive Safety Preference Dataset for Vision-Language Models
https://github.com/pikepokenew/ContraSafe-VL

>Logit-Conditioned Diffusion Decoding for Frozen Discrete-Token VLMs
https://github.com/jiwoohong93/dcdd_code

>SANEval: Open-Vocabulary Compositional Benchmarks with Failure-mode Diagnosis
https://github.com/nielseni6/saneval

>Fizgig 9.6: train LoRAs on MiniMax H3 Image
https://github.com/shootthesound/Fizgig/releases/tag/v9.6.0

>ComfyUI v0.45.0
github.com/Comfy-Org/ComfyUI/releases/tag/v0.45.0

>H3-FaceRefine-Plus - 2x face details
https://github.com/AnniCodex/ComfyUI-H3-FaceRefine-Plus

>MiniMax-H3-Image for Forge Neo
https://github.com/aoleg/pruna-P-H3-Image-forge-neo

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

File: debo_news_00039_.png (2.46 MB, 1792x1312)

2.46 MB PNG

>>110300168
>mfw REAL Research news

11/11/2026

>When to Start Denoising? Adaptive Denoising Entry for Generative Video Compression
https://arxiv.org/abs/2611.13620

>LongSCOP: Semantically Consistent Long Video Outpainting
https://arxiv.org/abs/2611.12445

>MPEG-Diffusion: Mixed-Resolution Video Generation
https://arxiv.org/abs/2611.12069

>Continuous Angular Control in Text-to-Image Generation via Geometric Token Encoding
https://arxiv.org/abs/2611.13172

>PhyRubric: Rubric-Grounded Diagnostic Evaluation of Physical Commonsense in Text-to-Video Generation
https://arxiv.org/abs/2611.13576

>Few-Step Streaming Video Generation via Condition Shifting
https://arxiv.org/abs/2611.14033

>Lipschitz-Controlled Flow Matching for Robust Image Restoration
https://arxiv.org/abs/2611.13995

>Shaping Useful Noise: Energy Distributions Predict Visual Pretraining Quality
https://arxiv.org/abs/2611.11962

>Direct Trajectory Optimization for Inference-Time Alignment of Text-to-Image Diffusion Models
https://arxiv.org/abs/2611.11821

>DUO: Dual Readouts of Diffusion Models for OOD Detection
https://arxiv.org/abs/2611.13304

>Four-Step Text-Bearing Image Generation via Distillation with Local Response Regularization and Teacher-Guided Training
https://arxiv.org/abs/2611.12042

>Just Image Transformers with Late-Stage Pixel Refinement
https://arxiv.org/abs/2611.13120

>Where Do Vision-Language Models Get Left and Right? Position-Code Counterfactuals in Qwen2.5-VL, Qwen3-VL and InternVL3
https://arxiv.org/abs/2611.14060

>EditJudge-Bench: Auditing VLM Image-Edit Judges with Synthetic Ground Truth
https://jenspars.github.io/EditJudge-Bench

>Advancing a Global Framework for AI Safety and Governance for the Well-being of Humanity
https://www.wicinternet.org/2025-11/07/c_1137524.htm

>What Social Media Teaches Us About Accumulative AI Risks
https://arxiv.org/abs/2610.24006

>Self-Distilling Generative Adversarial Networks
https://arxiv.org/abs/2605.08577

>>110301051
vibecode your own front end and make it look and operate however you want

>>110302295
oi m8 do you have a loicense for that?

File: 2026-11-11-134006_500x428(...).png (15 KB, 500x428)

15 KB PNG

*gasp

did anyone try that PAN2 video model?

>>110302370
this is what I did, its not perfect, but it runs the way I want/expect it to, as opposed to unintuitive comfy mess

>>110302476
killing myself

File: 1777258730717196.jpg (1.11 MB, 1408x2128)

1.11 MB JPG

{"inputs": {"lora_name": "metalheart-step00010080-a5659f-state.safetensors", "strength_model": 1.0, "model": ["1", 0]}, "class_type": "LoraLoaderModelOnly"}, "44": {"inputs": {"lora_name": "dodleyz-step00011270-ea99e3-state.safetensors", "strength_model": 1.0, "model": ["43", 0]}, "class_type": "LoraLoaderModelOnly"}, "45": {"inputs": {"lora_name": "shejanto-step00003480-41553f-state.safetensors", "strength_model": 1.0, "model": ["44", 0]}, "class_type": "LoraLoaderModelOnly"}

>>110302476
they're 100% trying to find a way to fuck us scrapers

>>110302605
No, they were down because they got hacked by some guy submitting a pull request that re-enabled flash uploads which he then used to uploaded a malicious flash worm.

>>110302668
insanely based
fuck danbooru they deserve it

File: 2026-11-11-140435_614x727(...).png (48 KB, 614x727)

48 KB PNG

>>110302476
ohno

>>110302683
They now let you use 3 tags for search instead of 2 tho.

Krea2fags on /b/ are seething about MiniMax Image being the GOAT kek

>>110302711
>Krea2fags on /b/
why are they not using the literal cheese pizza machine known as moss

I've been out of the loop for a while.
I'm still using qwen to edit images. Is there a new model for image editing that doesn't mess up the original pic?

>>110302752
Flux 3 Klein, it retains the colors and identity way better than previous models
problem is that it sucks apart from that

>>110302668
kekd
>>110302752
>doesn't mess up the original pic?
pixelspace editing is this coming year

>>110301215
maybe can you provide a toml for this? I did your past screenie with your settings and it's not a good output so it's likely I mistypedsome things, it's even missing some things like warmup steps.

i hate myself for buying an amd card years back when i could just have bought an nvidia
at least im on loonix and i get it somewhat to work but feel free to point at me and laugh

>>110303062
it's not that bad anymore, AMD sorta works these days

>>110303062
just use a sd.cpp based ui retard
all your pain will go away

>>110302942
I'm almost positive the toml was also present in the other screenshots I gave you.

>>110303204
i'll take a look, it's kinda hard to find with so many cross threads, i'd also like the toml for images as i'm trying to add random crop padding 0.05 on it.

>>110303221
>the toml for images as i'm trying to add random crop padding 0.05 on it.
I didn't do anything fancy like that.

>>110303228
>>110292167
checking the metadata, it has more of the settings than the pic sent, even the LR went from 0.00015 to 0.0001 with warmup stable decay.

>>110303249

  --attn_mode=flash \
  --network_module=networks.lora_generic \
  --network_dim=32 \
  --network_alpha=32 \
  --network_train_unet_only \
  --learning_rate=1e-4 \
  --optimizer_type=AdamW8bit \
  --lr_scheduler=warmup_stable_decay \
  --lr_warmup_steps=0.1 \
  --lr_decay_steps=0.2 \
  --timestep_sampling=truncated_logit_normal \
  --logit_mean=-1.09 \
  --logit_std=0.95 \
  --discrete_flow_shift=1.0 \
  --max_train_epochs=70 \
  --save_every_n_epochs=7 \
  --mixed_precision=bf16 \

You're right, the screenshot wasn't the same settings as the bake I ended up using.

>>110303319
kek

File: mossy.png (15 KB, 303x277)

15 KB PNG

>>110303283
thanks for this, that's all the changes to the original pic? i'm also curious on your .txt tagging, is it purely NL, and does it include linebreaks? i'm not fluent nor attempeted a NL style of tagging but it's worth a shot. i'll try baking this again overnight and see if it's a huge improvement.

File: 1793851830181340.mp4 (3.76 MB, 1664x1500)

3.76 MB MP4

Which ai do I need to use to be able to accurately track a person’s facial expressions and put it on another character’s face like this?

>>110303319
recently one man threw away billions on a mediocre woman. the hottest woman alive must be unfathomable value. and yet...the choice is clear

>>110303343
minimax h3 no i will not spoon feed you further

>>110303337
>all the changes to the original pic?
as far as I know yes.
my .txt files are pure NL, I don't think I have any line breaks but technically moss should handle them just fine and they are preferred when your captions contain rendered text.

Like,
A large title that reads "First line
Second line" is visible at the top.

>>110303376
i'll work on the tags later, thanks. it's honestly more work for tagging than booru style but hopefully it'll pay off especially when future official finetunes show up,
I read the moss documentation (lacks on what each thing does as I tried came and never worked.) it has something called "multiple candidates", a moss exclusive where it can have both NL and booru in a single .txt file but doesn't have an example on how to do it as further stuff beyond it make zero sense to me, hell even the part where it said to use \ for linux, ^ for windows gave heavy errors until the true fix was ` to finally make it readable with copy-paste abilities.

Gonna train a lora of my GFs favorite artist.

verdict on mememax h3 image?
KEEP IT OR SNEED IT?

>>110303712
Feed it

File: 2026-11-11-162234_194x119(...).png (23 KB, 194x119)

23 KB PNG

>>110303712

>>110303709
>not training a lora on your gf
>look hon this is how you'd look like sucking BBC wanna try ;)

>>110303709
cute :3 show her tits?

>>110303712
Has a lot of potential but is gimped to hell and back by prompting methodology

I have been seated for

hm.... thinking of what to prompt right now...

>>110304072
copy paste the whole thread into an llm, and tell it to make a prompt based on the contents and see what happens

>PAN 2
Yay or Nay?

>>110302752
FIBO Edit (non-Turbo) is SOTA (better than Qwen 2.1 Edit), F3K is a meme not sure why that other anon even mentioned it

>>110301014
this did not age well

>>110304149
Actually, it should be mentioned that FIBO Edit is better at edits than the latest GPT Image and Nano Banana models as well.

>>110304095
i did and it told me to kms :/

>>110304172
cutie
imagine him hugging a muscle mommy

>>110304170
i mean, you could still gen that

File: 80696.png (915 KB, 768x768)

915 KB PNG

Am I doing something wrong? I am trying to train a lora on moss using ai-toolkit using a 3060 trying to train a high detail 3D lora using 110 images (picrel was trained on Civitai) but the whole night and only 50 step progressed I'm getting like a step each 20 minutes... I'm seeing the gpu only using about 50 watts instead of the 170 peak it has, did I do something wrong?

>>110304149
I mentioned it because I just tried it out today

>>110302974
What did he do btw? The rentries are complete schizo babble

>>110304321
he was just a lying retarded attention whore who doxxed himself

>>110304321
hi ani kill yourself

>>110304237
Your card is fucked. Sorry, anon.

File: 5929.png (892 KB, 768x768)

892 KB PNG

>>110304391
I can gen and play games with it just fine, I am just doing something wrong which I don't know yet.

>>110304438
That image is fried in at least 3 different ways
Lower you tone when you speak to me

>>110304344
>>110304360
Can either of you show proof?

File: 2521.jpg (51 KB, 768x496)

51 KB JPG

>>110304523
The only thing that's fried is the dope you being smoking.