12 GB card numbers: int8 DiT + full 32B text encoder (RTX 4070 SUPER)
We tested the int8 files from this repo on an RTX 4070 SUPER 12 GB with 96 GB RAM: minimax_h3_image_pruned_int8_convrot, the full text encoder qwen3vl_32b_minimax_h3_int8_convrot and the fp16 VAE. ComfyUI v0.45.2 in a separate copy, the image_minimax_h3_image_t2i template with the prompt expander removed, 6 steps, CFG 1, euler / simple.
Speed at 2688x1536 (4.1 MP): 23 s for the first image with a cold load, then 11-12 s per image with a new prompt and 8 s when the prompt repeats (encoder output cached), about 1.1 s per step. With one reference image at 4.2 MP: 12 s warm, about 1.7 s per step. Dynamic VRAM streams the 21 GB DiT from system RAM, around 41 GB of RAM was in use. We never needed the w4a8 file. With only one text encoder loaded we did not hit the encoder VRAM issue (ComfyUI #17649).
Quality, side by side with Qwen Image 2.1 on the same 8 scenes x 2 seeds plus 4 fantasy illustrations: H3-Image follows the prompt more precisely and puts more small detail into the frame at 4 MP (object placement, embroidery, star fields, fine background). What we liked less: lighting is flatter and brighter with less drama, skin and surface texture are simpler than Qwen, and it does not hold a painterly style - with a comic-style reference it goes to the clean cel shading of the reference. Photo scenes look plausible. Overall it reads more like a classic illustration than Qwen.
The lich scene shows the prompt following best: on both seeds H3-Image puts the lance into the lich and bends the wall behind them, our Qwen render misses the lich. The market pair is the same prompt and seed.
Follow-up at ~16 MP, since the card says the model works best there. Same setup as above, only 2x per side: 5376x3072 (16.5 MP) for the photo scenes and 3712x4480 (16.6 MP) for the fantasy scene with a reference image. It runs fine on the 12 GB card: about 41 s per image warm (49 s for the first one after start) and 44 s with one reference image (60 s for the first). That is roughly 4x the 4 MP time for 4x the pixels.
Texture is clearly better at 16 MP. Wrinkles, fine skin detail and single hair and beard strands show up where 4 MP gave smooth skin. On older faces it sometimes goes too crisp, almost etched, and stray hairs can come out as thin bright lines across the face. We also removed "natural skin texture" from the prompt and the result barely changed, so here it is the resolution that matters, not the wording.
The same seed gives a different composition at 16 MP. In the fantasy scene seed 7201 still puts the lance into the lich, while seed 5171 fires it out of the frame, away from the lich. The style still follows the comic reference.
The crops below come from the same prompt and seed at both resolutions, the crop size in pixels is under each panel.
All sheets, including the market woman crop: https://huggingface.co/datasets/LIOSAM007/h3-image-12gb-tests