Skip to main content
Back

Veda Slashes MiniMax H3 Video Generation by 2.9x on Consumer GPUs

A distilled predictor skips 90% of attention tiles in MiniMax H3, delivering up to 2.9x faster text-to-audio-video generation in ComfyUI without touching model weights.

  • Veda team ships official ComfyUI node for MiniMax H3 with ~2.9x end-to-end video speedup.
  • A distilled predictor keeps only the top ~10% of attention tiles and skips the rest.
  • 7.1x attention speedup on RTX 5070, up to 6.8x on RTX PRO 6000 for 14 second clips.
  • Not a LoRA or quantization: weights are untouched, compatible with any LoRA or finetune.
  • Install via comfy node install veda-sparse-attention, needs ComfyUI 0.38.0+, Triton or MLX.
  • 275 MB fp8 predictor on Hugging Face works across T2VA, FL2VA, R2VA and any step count.

MiniMax H3 generates audio and video together, but its attention workload grows quickly with clip length. Veda’s new ComfyUI custom node predicts which sections of the attention map matter, skips most eligible tiles, and reports a 2.9× per-step speedup on an RTX 5070. Faster iteration on consumer GPUs comes without modifying H3’s backbone weights or installed LoRAs.

The release combines a team-maintained ComfyUI node with a 275 MB predictor checkpoint. Researchers from ByteDance, the University of Hong Kong, and the University of Science and Technology of China presented the underlying sparse-attention method at ICML 2026.

How Veda prunes the attention map

Standard attention compares every query token with every key token. Video adds tokens across space and time, causing the number of comparisons to grow roughly with the square of the sequence length. Veda divides that work into tiles, predicts the tiles most likely to affect the result, and computes roughly the top 10% of eligible video-attention tiles.

During predictor training, max-pooled full attention creates tile-level targets. At inference, the smaller predictor reconstructs those targets and drives the sparse kernel through the following path:

  1. Estimate tile importance. A Triplet Pooling estimator uses per-head query and key projections to score regions of the attention map.
  2. Select work per head. Head-aware top-k selection converts those scores into a tile mask.
  3. Load selected values. The kernel fetches only the chosen key and value tiles, avoiding construction of the full attention map.

Each layer and attention head can use different tile geometry, allowing Veda to handle spatial and temporal heads within a fixed hardware budget. H3’s backbone weights, fine-tuned variants, LoRAs, and learned style remain unchanged because the selector operates only on attention execution.