| # R2VA preview deployment notes | |
| - Run the README commands from a Miowtion checkout with Ref2VA encoding and | |
| independent reference/current tile budgets. See | |
| `docs/features/r2va.md` in that checkout for the input and cache contract. | |
| - Load the MiniMax-H3 Ref2VA backbone and Turbo 8-step LoRA separately. | |
| - This safetensors file contains the Veda predictor and tile plans. | |
| FP8 weights are dequantized to BF16 when loaded. | |
| - The example uses 32 reference tiles and 32 current-video tiles; | |
| text, VLM conditioning and audio remain fully attended. | |
| - Run `scripts/prepare_official_r2va_demo.py` from the Miowtion checkout | |
| to prepare the official H3 example. | |
| It requires Miowtion, Python, ffmpeg and ffprobe, but no API key or GPU. | |
| The official PE is preserved. The video soundtrack is exposed as | |
| `<Audio 1>` and the external voice reference as `<Audio 2>`. | |
| - For a private model repository, authenticate with `hf auth login` before | |
| downloading the predictor. | |