Paper · Project Page · Predictor · Training Code · 中文说明

Introduction

Veda is a learned sparse-attention method for video diffusion models. Attention dominates the cost of a long clip, and a small distilled predictor can say in advance which tiles of the attention map carry the result. Veda computes roughly the top 10% of them and skips the rest.

This repository packages Veda as a custom node for ComfyUI's native MiniMax-H3 (T2VA, FL2VA and R2VA): one node, MODEL in and MODEL out.

Highlights

• Faster inference. 2.9x end to end and 7.1x on attention alone against ComfyUI's default attention, measured on an RTX 5070. The gain grows with clip length.
• Plug-and-play. An attention override rather than a weight patch, so Turbo and style LoRAs, fine-tuned and quantized H3 checkpoints, first/last-frame and reference conditioning, and other block patches keep working.
• Safe by default. Every kernel passes a self-test on the GPU before it is used. Anything Veda cannot handle runs the model's own attention and says so on the node, rather than producing a broken render.
• Nothing extra to install. The sparse kernel ships with the node; installing from the Comfy Registry pulls in Triton (NVIDIA, SM80 and newer) or MLX (Apple silicon).

Installation

Requires ComfyUI >= 0.38.0. In ComfyUI Manager, search "Veda". Or:

comfy node install veda-sparse-attention

Or clone this repository into ComfyUI/custom_nodes/ and restart ComfyUI.

Usage

Open Workflow -> Browse Templates -> Veda-on-ComfyUI and pick "Veda MiniMax H3 T2VA" or "Veda MiniMax H3 R2VA". The missing-model dialog offers the predictor. Write a prompt and run.

To add Veda to an existing H3 workflow, put Veda Sparse Attention (MiniMax H3) on the MODEL wire after the model and any LoRA loaders, last before the guider or sampler. Select it and press Ctrl+B to bypass it and render the same seed with full attention.

Do not combine it with ComfyUI's own "Model Sparse Attention" node on H3: that one replaces the attention blocks outright, so Veda would never be called. The node says so when it sees both.

Other attention nodes can stay. Veda runs on top of "Patch Sage Attention KJ" (the calls it does not take go to sage), and KJNodes' "MiniMax H3 Low VRAM Attention" keeps working. A node that swaps out the whole H3 attention, such as "MiniMax H3 Mem Eff Sage Attention Patch", must come before Veda: Veda then runs the sparse layers and leaves that node the full-attention ones, and says so on its node. Placed after Veda it would hide every call from Veda; the node warns about that too.

Predictor

The node does not download anything itself. The predictor (275 MB) reaches models/veda one of two ways:

• From a template. Open one of the Veda templates and ComfyUI offers the predictor in its missing-models dialog; one click fetches it.
• By hand. Download minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors from the predictor repository into ComfyUI/models/veda/ and restart ComfyUI.

Any .safetensors in that folder appears in the predictor list, so a predictor trained elsewhere is selected the same way. If the selected file is not there, the node says so and prints the URL rather than reaching out on its own.

Node text

Veda done · Triton INT8 (SM120)
Video: 1344x768 · 5.2 s
Attention computed: 10.9% of full attention (89.1% skipped)

Before sampling the node names the kernel it will use and the sparsity; while sampling, the video size and the trained tile plan it matched; at the end, how much of full attention was actually computed. A line starting Veda off gives the reason Veda is not running, and nearest trained size: ... on the tile plan line means the video is outside the trained set. verbose adds per-phase timing, call counts and predictor details.

Warnings and errors come in English and Chinese, one after the other; the node's name and tooltips follow ComfyUI's interface language.

Options

Only model and predictor are visible. The rest are advanced inputs (click "show advanced inputs") and default to the trained values.

The released predictor was trained for 1344x768, 768x1344, 768x768 and 1024x768 at 5 / 10 / 14 s with the 8-step Turbo LoRA. Other sizes fall back to the tile plan of the nearest aspect ratio and duration. Other sizes, other step counts and R2VA / FL2VA references all work, but are outside the training data, so compare them against full attention.

Performance

T2VA at 1344x768, 124 frames (5.2 s), 8-step Turbo LoRA, 90% sparsity, on an RTX 5070 12 GB under Windows 11.

One attention layer at 104k tokens and 90% sparsity takes 528 ms against 11.8 s for full attention. Veda's INT8 arithmetic is the same code ComfyUI runs behind --use-sage-attention, so the quality is what that path gives, with sparsity on top. What remains per step is MLP and weight movement, not attention.

Hardware

One Triton kernel covers every NVIDIA GPU from SM80 on, Windows and Linux alike. Older cards, ROCm and CPU have no kernel: the node says so and the model runs its own attention.

Per-architecture notes and the full measurement log are in docs/hardware.md.

License

Code: MIT. The INT8 kernel's arithmetic is derived from SageAttention v1 (BSD-3-Clause). The predictor inherits the MiniMax H3 Community License from its base model. See NOTICE.md.

Citation

@inproceedings{han2026veda,
  title={Veda: Scalable Video Diffusion via Distilled Sparse Attention},
  author={Han, Shihao and Yang, Hao and Hu, Xinting and Mei, Xiaofeng
          and Jiang, Yi and Qi, Xiaojuan},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2026}
}