Skip to content

feat: add Tiny VAE previews for MiniMax H3 - #2124

Open
Pheonotype wants to merge 2 commits into
deepbeepmeep:mainfrom
Pheonotype:main
Open

Pheonotype wants to merge 2 commits into
deepbeepmeep:mainfrom
Pheonotype:main

Conversation

@Pheonotype

Copy link
Copy Markdown

Adds an optional true-RGB sampling preview for MiniMax H3 using the 24-channel H3 Tiny VAE decoder.

Changes

  • Adds automatic download and loading of the ~40 MB H3 Tiny VAE decoder.
  • Adds MiniMax H3 Tiny VAE Preview under Advanced Mode → Misc.
    • Off: existing latent preview
    • CPU: low-VRAM Tiny VAE preview
    • GPU: faster Tiny VAE preview when sufficient VRAM is available
  • Uses the TAEHV sequential temporal decode path to avoid allocating the entire video preview at once.
  • Correctly converts Tiny VAE output from [-1, 1] to 8-bit RGB.
  • Falls back to the standard latent preview if the Tiny VAE cannot load or decode.
  • Documents the feature in the MiniMax H3 handler help text.

Validation

  • Python syntax checks pass.
  • The original taeh3_decoder.safetensors loads with strict state-dict validation.
  • Decoder smoke test passes with the memory-efficient temporal path.

GPU memory/performance should be validated across H3 resolutions and VRAM profiles.

@GOvEy1nw

Copy link
Copy Markdown

Haha, I added the same req for LTX (to start) too, but yes this would be great. #2107

@Proximo027

Copy link
Copy Markdown

YES! First of all, thank you for this amazing program!! I will agree here as my 10 sec videos with MH3 and head ref take me about 18 minutes for a 10 sec video at 1280 x 704. The preview window does not cover the whole video. My very last video ended up with two people in it instead of one. It would have been so nice to see this early and cancel.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants