# H3 Image for Forge Neo — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/866436b681674057

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-11 00:36 ingested 2026-11-12 02:40

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xt3ovz/h3_image_for_forge_neo/)

## retained text



```
Basically, the extension: [https://github.com/aoleg/pruna-P-H3-Image-forge-neo](https://github.com/aoleg/pruna-P-H3-Image-forge-neo) \- supports MiniMax-H3-Image and PrunaAI P-H3-Image in Forge Neo.

Tips:

* ~~Heun / P-H3-Image / 8 Steps / CFG=1 is the golden combo. Euler / Simple / 6 Steps will be noisy. You will need to use P-H3-Image as a scheduler; everything else won't work; Euler or Heun is your choice. Also, use 8 steps, not 6 or 10.~~
* UPDATE: **LCM, Simple,** 6 steps looks better than Euler or Heun, and is faster. **Euler a** looks better than plain Euler even with Simple. Recommendation: use **LCM, Simple, 6 steps** or **Euler a**, H3-Image 8-step, 8 steps.
* You don't really need the 32B encoder (besides, the nvfp4 version won't work). You can use the much smaller P-H3-Image text encoder in INT8-Convrot with great result.
* Yes, editing works, too. Image restoration also.
* You can use unimaginably high resolutions without OOM, but you'll likely need to edit your ui-config.json to allow width and height values over 2048.
```



document fetched 2026-11-12 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xt3ovz/h3_image_for_forge_neo
