ComfyUI-SPEED node cuts Anima sampling time nearly in half
A new ComfyUI custom node implements SPEED (Spectral Progressive Diffusion) for Anima, dropping generation time from 26 to 14 seconds by starting at lower resolution and scaling up during denoising.
A ComfyUI user released ComfyUI-SPEED, a custom node that brings Spectral Progressive Diffusion sampling to Anima workflows. The node cuts generation time roughly in half — 14 seconds versus 26 seconds in the developer's tests — by starting diffusion at lower resolution and progressively increasing detail as the image forms. The implementation is available on GitHub.
SPEED (Spectral Progressive Diffusion) skips full high-resolution work in early denoising steps, where detail isn't yet visible. By upsampling gradually through the sampling process, the method reduces wasted compute without sacrificing final image quality. The developer describes the node as "vibecoded" — functional but not polished — and notes that official SPEED code hasn't been released yet, so this is a community interpretation of the technique described at howardxiao.ca/speed.
What stands out
- 01Speed gains are real but hardware-dependent. The developer's test showed a 26-second baseline dropping to 14 seconds with SPEED enabled, both using the same seed. That's a 46% reduction, though results will vary by GPU and prompt complexity.
- 02Artifacts can appear. The node can produce visual drift or upsampling artifacts on some outputs. The developer flags this as a known issue, likely tied to how the progressive resolution scaling is handled.
- 03torch.compile doesn't help here. Enabling PyTorch's JIT compilation actually slowed sampling in testing, contrary to what you'd expect with most custom nodes. The developer recommends leaving it off.
- 04