← sources

MiniMax-H3-Image at 16 MP, 6 steps, ~20 s per image

source: reddit (reddit:StableDiffusion)published 2026-11-11 16:02ingested 2026-11-12 02:40

retained text

Single seed, no upscale, no inpainting. Hand-written prompts, no prompt expander. \~20 s on an RTX PRO 6000, \~22 s on a 4090 with the int8 weights

Settings  
\- Model: minimax\_h3\_image\_pruned\_bf16 (Comfy-Org pruned DiT)  
\- Text encoder: qwen3vl\_32b\_minimax\_h3\_bf16, CLIPLoader type "minimax"  
\- VAE: minimax\_h3\_image\_vae\_fp16  
\- Resolution: 4160x4000 (\~16.6 MP), set on MiniMaxH3ImageConditioning  
\- Sampler: euler / simple, 6 steps, CFG 1, negative = ConditioningZeroOut  
\- No prompt expander, prompts written by hand (4-5 sentences each)  
\- ComfyUI 0.45.0

document fetched 2026-11-12 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xtlj1t/minimaxh3image_at_16_mp_6_steps_20_s_per_image