P-H3-Image vs official H3-Image on a 12 GB card (int8, same prompts)

#12
by LIOSAM007 - opened

We ran the int8 set from this repo (p-h3-image_int8_convrot plus the int8 text encoder and VAE) on an RTX 4070 SUPER 12 GB with 96 GB RAM, ComfyUI v0.45.2, the core graph from the card: 8 steps, euler, your ManualSigmas, no CFG. At the recommended resolution (8.3 MP) it takes 33-40 s per image, about 3.3 s per step. For reference, official MiniMax-H3-Image int8 at 4.1 MP takes 8-12 s on the same card.

We rendered 8 scenes x 2 seeds plus 4 fantasy scenes with a reference image x 2 seeds and compared them side by side with official H3-Image on the same prompts. P-H3 gave nicer atmosphere and light (fog, light rays, night sky) and sometimes a richer crowd. We did not get the skin texture gain we hoped for: in photo scenes the skin came out smoother and more retouched than with the official model, and with a comic-style reference the result had a glossy airbrushed look. Following the card, we reran the fantasy scenes with a style-first prompt and without words like cinematic and highly detailed. The images barely changed, so the style seems to come from the reference rather than the prompt.

Same prompt, reference and seeds. In the lich scene P-H3 keeps the lich in the background on both seeds, and on seed 5171 the lance goes toward the camera. The face crop is the same area of the spice market frame from each model.

lich: H3-Image vs P-H3-Image

face crop: H3-Image vs P-H3-Image

For our use we stay with official H3-Image and Qwen Image 2.1, but the core graph worked out of the box, thanks for that.

I also tested it. I like the P-h3 results better.
Also, unlike the official release, p-h3 preserves faces and works better with references.

I think Pruna is much better for my works than the official Minimax release!

Also, I’d like to say that with p-h3 you can use the new minimax-h3-image_vae and new qwen-text-encoder 32B. They are fully compatible and yield interesting results.

I tested both for edit and Pruna results are much better at reference following. I make character datasets and official implementation in Comfy changes appearance and body proportions dramatically. Official node also hardcode Vision and VAE-ref to small resolutions. Increasing both with custom node (VAE-ref to 4096) returns reference following, but also makes it run much slower.
I stick with P-h3 version for my case.

Sign up or log in to comment