ComfyUI-SeaCache

ComfyUI-SeaCache

An independent ComfyUI node adaptation of SeaCache, using its scheduler-aligned spectral (SEA) filter to decide when a native DiT can reuse its previous residual.

Supported native ComfyUI model families are FLUX, HunyuanVideo, Wan 2.1, and MiniMax H3. The first three are the families published in the upstream SeaCache repository; H3 is an implementation-specific extension for ComfyUI's native joint audio/video DiT. auto detects the loaded model; connect SeaCache immediately after the diffusion-model loader and after any LoRA/model patch nodes.

Install

Copy this folder to ComfyUI/custom_nodes/ComfyUI-SeaCache, then restart ComfyUI. It uses only PyTorch and ComfyUI, so there is no requirements.txt.

Settings

rel_l1_thresh controls the speed/quality trade-off. 0 bypasses the node. Values around 0.10 to 0.30 are a sensible starting range, but the right setting depends on model, resolution, sampler, and prompt. cache_device=cuda is fastest; use cpu only when VRAM pressure matters more than speed.

start_percent and end_percent define the sampling window in which caching is active. First cacheable step always runs in full, and the cache is reset for every sampling run.

For MiniMax H3, h3_metric_stride spatially subsamples the video tokens before its SEA FFT; the actual diffusion output is never downsampled. Start at 4 for high-resolution work, reducing to 2 only if there is enough VRAM. H3's audio change is also tracked, so video reuse cannot freeze a changing audio trajectory.

Caveats

• This targets current native ComfyUI DiTs. Diffusers-only loaders, wrappers with their own denoising loop, VACE/S2V Wan variants, and FLUX Kontext's special per-token timestep path are deliberately left on the native forward path.
• Do not stack it with EasyCache, TeaCache, or another whole-DiT caching node on the same model.
• SeaCache upstream currently has no repository license file. This package is a local adaptation that reimplements the published SEA gating method; confirm upstream licensing before redistributing it.

License

This ComfyUI node adaptation is licensed under GPL-3.0-only. SeaCache itself is separately maintained by its upstream authors; its licensing status does not change through this package.

Attribution

SeaCache: Chung et al., SeaCache: Spectral-Evolution-Aware Cache for Accelerating Diffusion Models (CVPR 2026). The node integrates with ComfyUI's native model interfaces and follows the MODEL -> MODEL node pattern used by ComfyUI caching nodes.