# Minimax H3 BF16 and face refinement. — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/acdbb233eb3773f1

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-28 04:37 ingested 2026-10-29 02:40

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xher16/minimax_h3_bf16_and_face_refinement/)

## retained text



```
[Face Refinement](https://preview.redd.it/ceu0ju5dz4yh1.jpeg?width=1280&format=pjpg&auto=webp&s=c5574d4c7a5f7b91a0dc81dfc9611230dfd9a640)

The minimax H3 Hybrid BF16 model, even pruned, is over 40GB. I don't have A/B testing of BF16 vs INT8 because the video/composition is ultimately different even with the same seed. However, subjectively, and I think it is becoming more well known that BF16 models while bigger, larger and slower, are superior in terms of motion, face, and image quality.

  
Over lots of 1girl rendering, I find that BF16 can achieve superior faces (especially with 50 steps, which was the case here). However, face refinement is still necessary and worthwhile as demonstrated.
```



document fetched 2026-10-29 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1xher16/minimax_h3_bf16_and_face_refinement
