Support Tiny VAE (taeh3.safetensors) for high-quality latent previews using MiniMax H3
- Dominant language
- Python
- Stars
- 133k
- Forks
- 15.7k
- Avg merge
- 1d 7h
- Merged PRs (30d)
- 158
Description
### Feature Idea
Add support for the MiniMax H3 Tiny VAE (taeh3.safetensors) to enable high-quality previews of MiniMax H3 latents during sampling.
Currently, previewing MiniMax H3 latents without the Tiny VAE results in poor-quality or unusable previews. The Tiny VAE provides a lightweight way to decode intermediate latents into meaningful video previews.
**Expected behavior:**
Detect/load taeh3.safetensors from ComfyUI/models/vae_approx
Display substantially higher-quality previews during sampling.
taeh3.safetensors link:
https://github.com/madebyollin/taehv/blob/main/safetensors/taeh3.safetensors
### Existing Solutions
_No response_
### Other
_No response_
Contributor guide
Research direction
Start by tracing how ComfyUI discovers and loads files in ComfyUI/models/vae_approx, using the existing approximate-VAE preview handling as the comparison point. Confirm that taeh3.safetensors is detected and loaded for MiniMax H3 latents, then verify that sampling previews are substantially higher quality than without it.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 52/100