Text-to-Image
Diffusion Single File
comfyui

Realisem quality in horrible

#5
by agapes - opened

i checked with lot of different settings and resolution but couldnt get anything good out of it
i checked 4-16MP and 6-30 step non of them even close to Krea2
maybe on cartoon or Anime fine but photo realisim forget about it

try using er_sde and sgm_uniform

yes same i am using bf16 but same not realistic

I've got this good result, without changing anything from the workflow
minimax_h3_image_t2i_00001_kvhji_1794367788

https://huggingface.co/MiniMaxAI/MiniMax-H3-Image/discussions/9

It's clear that this model performs much better. See the comparison tests below.

what i noticed and what i fixed was my default workflow and prompting was based on the minimax video model that i used for images using a custom workflow
it had a lot of realistic skin textures and skin pores in the prompt

the model already produces realistic result so removing those added on realism lines helped a lot

I don't know if this is a solution; I'm not an expert in AI models, but perhaps it could help. If you use minimax_h3_video_vae, the quality is better, with an image size of exactly 4 MP. Smaller image sizes always result in poor quality.

Try swapping the sampler to lcm (scheduler simple, 6 steps, CFG 1, denoise 1.0) before touching anything else. In our testing that single change eliminated the texture problems this thread describes β€” the mottled, coarse skin, the chalky fuzz on lit edges, the harsh microcontrast.

On the "why": the original version of this comment claimed the distilled heads are consistency-style endpoint predictors. That was wrong, and the correction is ours to make: per the PDD paper (Parallel Decoding Distillation, arXiv 2607.26004), the fused heads predict the interval mean velocity over each of the 6 grid steps. The honest state of our knowledge is that lcm/simple/6 renders clean here, reproducibly, and the bundled template's euler configuration did not in any arm we tested. We won't replace one tidy wrong mechanism story with another β€” treat the recommendation as empirical.

The pathology is sampler-side, not weights-side: we reproduced it unchanged across the bf16 and int8 DiT, both text encoders, and 4 MP through 16 MP β€” and the image VAE round-trips cleanly (~32 dB PSNR on a straight encode/decode, no tiling artifacts), so it isn't the VAE either. With lcm/simple/6, the same seeds render clean at 16 MP.

The bundled template defaults to euler, which is likely why first impressions here ran so rough. Hope it saves someone the hunt.

β€” Jingoro Ikeda, Goldfish Games

Edited 2026-11-21: corrected the mechanism paragraph β€” the original endpoint-predictor explanation was wrong; the recommendation is unchanged.

try SamplerLCM , s_noise = 1.10 ,s_noise_end =0.9 , noise_clip_std = 0
scheduler : simple ,steps : 6

I noticed that pdd_fused_heads.safetensors is also provided in the official repository.
Could this file be related to the issue? And how should we use it within ComfyUI?

Sign up or log in to comment