Warning: This repository is "vibecoded" — it may contain experimental code, unconventional styles, or project-specific shortcuts.

This repository provides a ComfyUI custom node that integrates the SPEED (Spectral Progressive Diffusion) idea for faster sampling. SPEED is a method for progressively growing image resolution during diffusion denoising to reduce computation while preserving visual quality. (only tested for anima model)

Key references

• Original project page: https://howardxiao.ca/speed/
• Paper (PDF): https://howardxiao.ca/speed/paper/paper.pdf
• arXiv: https://arxiv.org/abs/2605.18736

Summary

Workflow

Speed comparison (Anima, using the current default input config)

SPEED sampler (this node)	Baseline (euler sampler)
base_sampler=euler start_scale=0.50 mid_scale=0.75 transition_1=0.80 transition_2=0.60 taper=8 14.55s 1.82× faster	26.51s 1.00×

Notes

• Artifacts: This implementation can produce visible artifacts on some outputs; results may vary by model and prompt. Inspect the example images above for a representative comparison. taper (default 8) crossfades the DCT seam to reduce ringing at transitions; set taper=0 for the original hard-truncation behaviour.
• Torch compile: Compiling with torch.compile did not improve performance for this implementation and in our tests made sampling slower than running without it. It may be possible for others to make the node work with torch.compile, but this remains a known / open issue.
• Spectral Progressive Diffusion (SPEED) progressively increases resolution and injects higher-frequency components along the denoising trajectory, enabling training-free acceleration and a light fine-tuning recipe.
• Personally i recommend using transition_1 value of 0.8 and transition_2 value of 0.7. at 1.4x speed up for anima after many tries.

Usage

• Connect the Sampler SPEED (Spectral Progressive) output to SamplerCustomAdvanced like any other custom node in ComfyUI.
• Place this folder under your ComfyUI custom_nodes directory, then restart ComfyUI.

Inputs

• base_sampler: which underlying ODE solver to use. Any sampler in comfy.k_diffusion.sampling is supported (euler, euler_ancestral, heun, dpmpp_2m, dpmpp_3m_sde, er_sde, res_multistep, uni_pc, ipndm, …). euler is the original SPEED behaviour. Multistep solvers reset to first-order at each SPEED transition (solver state across a resolution change is invalid anyway).
• start_scale: the first resolution fraction. 0.5 means the sampler starts at half resolution, so it does the earliest denoising work on a smaller image.
• mid_scale: the second resolution fraction. 0.75 means it expands to 75% resolution before going to full size.
• transition_1: Controls the jump to mid-scale; lower values delay the expansion, forcing the model to spend more steps inside the low starting resolution.
• transition_2: Controls the jump to full-scale; setting this too low delays the final expansion, forcing the model to bake fine textures into a mid scale latent which stretches into blocky artifacts.
• taper: DCT-bin width of the cosine crossfade at the preserved↔noise boundary. 0 reproduces the original hard truncation (more ringing); higher values smooth the seam at the cost of slightly more low-pass. Default 8.

Credits & license

• This implementation links to and builds on the ideas from "Spectral Progressive Diffusion for Efficient Image and Video Generation" by Howard Xiao, Brian Chao, Lior Yariv, and Gordon Wetzstein. Please see the original project page and paper for full details, authorship, and license information.
• Thanks to acidmiku for contributing the pluggable sampler dispatch, cosine taper crossfade, and DCT matrix cache.

BibTeX

@article{xiao2026spectral,
  author    = {Xiao, Howard and Chao, Brian and Yariv, Lior and Wetzstein, Gordon},
  title     = {Spectral Progressive Diffusion for Efficient Image and Video Generation},
  year      = {2026},
}