Cache the conditioner client with functools.cache, tidy comments

Let gradio_client forward the caller's ZeroGPU token itself instead of pinning one by hand

Bill the conditioner call to the requesting user's own ZeroGPU token, never an org token

Pay for the conditioner call with the first ZeroGPU identity that can: the caller's token, this Space's HF_TOKEN, then none

Fall back to an anonymous conditioner call when the forwarded ZeroGPU token is refused

Forward the caller's ZeroGPU token to the conditioner so one request bills as one request

Load the AoTI packages from the public multimodalart/minimax-h3-aoti, move the upsample toggle under the prompt

Migrate to canonical diffusers PR 14371 + public MiniMaxAI/MiniMax-H3 weights, add prompt upsampling toggle

Update README.md

The keyframe half moved to multimodalart/minimax-h3

Update README.md

Point at the public multimodalart/qwen3vl-conditioner, called without a token

A real person as the subject reference (Pexels, Oliver Dohrn)

Real speech for the audio example: LibriSpeech dev-clean 1462-170145-0022

Dynamic GPU duration from the packed sequence, cached examples, torchaudio note

Rename to MiniMax-H3 Reference; describe the tab order and image slots

Describe the reference tabs

MiniMax-H3 ref2va, the denoising half of the split deployment

initial commit