Veda Sparse Attention for ComfyUI is a custom node that implements learned sparse attention for video diffusion models running on ComfyUI's MiniMax-H3 architecture.
The tool addresses the computational cost of attention in long video generation by using a small distilled predictor to identify which tiles of the attention map are essential to the output, then computing only those tiles while skipping the rest. This sparse-attention approach reduces attention computation to roughly the top ten percent of the full map. The node wraps this as a single custom node that sits on the model wire, accepting a model and returning an optimized model without requiring weight modifications.
Developers should adopt this tool if they are running MiniMax-H3 workflows in ComfyUI and want faster inference without sacrificing compatibility. The tool is designed as an attention override rather than a weight patch, which means existing LoRAs, fine-tuned checkpoints, quantized models, and other block patches continue to work unchanged. The node includes self-tests for every kernel before use and falls back to the model's native attention for any case it cannot handle, ensuring safe degradation rather than broken output. Installation requires ComfyUI version 0.38.0 or later and pulls in either Triton for NVIDIA GPUs or MLX for Apple silicon. The node warns when placed alongside incompatible attention nodes like ComfyUI's own "Model Sparse Attention" for H3, and it can coexist with other attention nodes such as "Patch Sage Attention KJ" or "MiniMax H3 Low VRAM Attention".
The project maintains active development with regular updates addressing edge cases and compatibility. The codebase shows careful attention to robustness, including built-in safety checks and clear user warnings for configuration conflicts. Documentation includes workflow templates and detailed guidance on node placement and interaction with other attention modifications.
