# Open-H3-Regenerate-2K to preserve draft pass details — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/6467a3478796b73a

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-02 21:47 ingested 2026-11-03 02:40

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xmal1l/openh3regenerate2k_to_preserve_draft_pass_details/)

## retained text



```
NVIDIA’s original Open-H3-Regenerate-2K recipe can struggle to preserve facial features and text established during the draft pass. ( see my previous post here: https://www.reddit.com/r/StableDiffusion/s/Pm77j7t6xr )

Valtz introduces an alternative refinement recipe that addresses these issues while retaining the same computational cost. The approach is not specific to Valtz and should be applicable to other frameworks as well.

We generated MiniMax H3 videos beyond 768p in two passes: a 768p draft, followed by NVIDIA’s Open-H3-Regenerate-2K refinement at the requested resolution.

The key difference is where the refinement begins on the noise schedule. Valtz offers two options:

\* Keep the Draft (Valtz’s default since October 31, 2026): Re-noises the upscaled draft to σ = 0.7, then denoises it in three steps (0.700 → 0.579 → 0.382 → 0). Roughly 30% of the draft signal survives, helping preserve facial features, lettering, and other details established in the first pass.

\* NVIDIA’s Recipe (the original release): Starts at σ = 0.909 and follows a three-step schedule (0.909 → 0.857 → 0.667 → 0). Only about 9% of the draft signal survives, giving the refinement pass much more freedom to redraw faces and text.

Both options use the same merged LoRA, take the same three denoising steps, and have essentially identical runtimes. The trade-off is straightforward: Keep the Draft improves consistency, but also preserves mistakes made during the draft pass.

The grain visible in some 2K outputs is associated with Sol Attention in the refinement pass under the Favor presets, rather than the choice of starting sigma.

The attached clips were generated at 1440p using the Keep the Draft recipe. The wonderful prompts are borrowed from the open VDN project’s page. For the photographer–tram–war scene, the draft pass was rendered using VDN, which proves to be a solid alternative to Sol Attention, particularly at this resolution and video length.
```



document fetched 2026-11-03 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xmal1l/openh3regenerate2k_to_preserve_draft_pass_details
