MiniMax H3-Image and P-H3-Image on a 12 GB card

Comparison sheets from our local test on 11.11.2026: RTX 4070 SUPER 12 GB, 96 GB RAM, ComfyUI v0.45.2.

• MiniMax-H3-Image int8 (Comfy-Org/MiniMax-H3-Image), full text encoder qwen3vl_32b_minimax_h3_int8_convrot, fp16 VAE, 6 steps, CFG 1, euler / simple, prompt expander removed. 4.1 MP: 23 s cold, 11-12 s warm.
• P-H3-Image int8 (PrunaAI/P-H3-Image-Comfy), core ComfyUI graph from the card, 8 steps, euler, ManualSigmas. 8.3 MP: 33-40 s.
• Qwen Image 2.1 int8 frames are from our regular graphs.

Files:

• h3image_lich_vs_qwen.jpg - fantasy scene with a reference image, Qwen Image 2.1 vs H3-Image, two seeds
• h3image_market_vs_qwen.jpg - spice market, same prompt and seed
• ph3_vs_h3image_lich.jpg - the same fantasy scene, H3-Image vs P-H3-Image, two seeds
• ph3_vs_h3image_skin_crop.jpg - face crop from the spice market frame, same area of each frame

Follow-up at ~16 MP (H3-Image int8, 5376x3072 and 3712x4480, about 41 s per image warm on the same card):

• h3image_4mp_vs_16mp_market_merchant.jpg, h3image_4mp_vs_16mp_market_woman.jpg, h3image_4mp_vs_16mp_kitchen.jpg - face crops, 4 MP vs 16 MP with and without 'natural skin texture' in the prompt
• h3image_4mp_vs_16mp_lich.jpg - the fantasy scene at 4 MP and 16 MP, two seeds

Discussions:

• https://huggingface.co/Comfy-Org/MiniMax-H3-Image/discussions/12
• https://huggingface.co/PrunaAI/P-H3-Image-Comfy/discussions/12