Veda R2VA for MiniMax-H3: reference-aware sparse attention, budgets and measured speedups

Veda-Sparse also publishes an R2VA predictor for MiniMax-H3 reference-to-audio-video generation, separate from the T2VA predictor. It was fine-tuned for 600 updates from the T2VA predictor on 5.17-second R2VA clips. It selects visual-reference tiles and generated-video tiles independently, with fixed budgets of 32 reference tiles and 32 current-video tiles per query tile and head; text, VLM conditioning and audio attention remain dense. It is not a LoRA and leaves the backbone untouched; the predictor is stored in FP8 and dequantized to BF16 at load. Reported speedups with the 8-step Turbo LoRA on one RTX PRO 6000 Blackwell (attention / end-to-end): 5.17 s clips at 16:9 2.42x / 1.69x, 9:16 2.65x / 1.67x, 1:1 2.06x / 1.37x, 4:3 2.19x / 1.54x; 16:9 at 10.1 s 2.84x / 2.05x and 14.4 s 2.90x / 2.20x. These are lower than the T2VA predictor's figures, and longer clips gain more. Samples are 1344x768, 124 frames. Setup (Miowtion main branch): git clone --recurse-submodules --branch main https://github.com/veda-sparse/Miowtion.git, pip install -e '.[gpu,encode]', download MiniMaxAI/MiniMax-H3 (model_index.json and Ref2VA/*), larryvrh/MiniMax-H3-Turbo-Lora minimax_h3_turbo_v4_step600_ema.safetensors and this repository. Then scripts/prepare_official_r2va_demo.py, scripts/encode_samples.py and scripts/generate.py --config configs/infer_r2va_preview.yaml --predictor <minimax_h3_r2va_veda_preview_fp8.safetensors> --attention dense veda (renders both for comparison). AGENTS.md has GPU-specific notes. Licence: inherits the MiniMax H3 Community License. Based on the Veda-Sparse/Minimax-H3-R2VA-Veda-Preview card pinned to revision 1598b407905b7c0af85aa0c13acdfa6f01ff2982 and observed on 2026-10-27; a preview release. Nothing was run here; the comparison videos, AGENTS.md and ComfyUI support for R2VA were not inspected. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/Veda-Sparse/Minimax-H3-R2VA-Veda-Preview/raw/1598b407905b7c0af85aa0c13acdfa6f01ff2982/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 734:1052, 1054:1366, 12283:12344, 12797:13020, 13418:13534; exact excerpt SHA256 b72470f22120df1f120616c6101e601d51fcde8cb77caebf32ca42bf4f82a9ff,e8ce28c243a4a0af637d162481063c6bfc7878356698f44b26e7d2c4ab4fc250,9047b5e6c753b1f6af0670f452b2bc6ffbb18846db5d20cee424b783cc792fdc,55"
    }
  ]
}