# A new Stills mode for MiniMax with a new Latent & VAE Decode node. — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/e1c3d3c527173528

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-09-25 22:06 ingested 2026-09-26 02:37

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1wq9dnt/a_new_stills_mode_for_minimax_with_a_new_latent/)

## retained text



```
8mp Example here where it sings more with less plastic and more skin detail when you go for the higher res: [https://www.reddit.com/r/StableDiffusion/comments/1wq9dnt/comment/pc2mg4c/?utm\_source=share&utm\_medium=web3x&utm\_name=web3xcss&utm\_term=1&utm\_content=share\_button](https://www.reddit.com/r/StableDiffusion/comments/1wq9dnt/comment/pc2mg4c/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button)

**3840x2176 still at 50 steps with no turbo lora takes 80 secs on my 5090 fyi if you want a benchmark.**

I dug deep into H3 Latents and VAE Decoding the last few days and this is the result.  Its not perfect, but its a big big improvement I feel.  Its FAST too **(single frame latent is MUCH faster than 5 frame)**, and as per the readme I use the turbo lora only on a low setting in the provided workflow.  But you can easily turn it off and increase steps.  Take note of the turbo lora used, and sampler/scheduler.  There may well be even better combos, but this one is a good one and took a fair bit of hunting to find!

**EDIT:**  There is also a Edit workflow included that demonstrates the model already has some edit abilities  (the 2.5mp upscale is intentional in the edit flow, it seems to work best at that mp)

[https://github.com/shootthesound/ComfyUI-Fizgig-H3-Still](https://github.com/shootthesound/ComfyUI-Fizgig-H3-Still)
```



document fetched 2026-09-26 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1wq9dnt/a_new_stills_mode_for_minimax_with_a_new_latent
