ruwwww/comfyui-speed

This project is an add-on for ComfyUI, a node-based tool people use to generate images from AI models. The add-on implements an idea called SPEED, which stands for Spectral Progressive Diffusion. The point of SPEED is to make image generation faster by starting at a lower resolution and growing the image during the generation process instead of working at full size from the beginning. The author says the implementation has only been tested on a model called Anima. The README marks the project as vibecoded, which the author uses as a warning that the code is experimental and may have rough edges. It links to the original research project page, a paper PDF, and an arXiv listing for the underlying method. A comparison table shows the SPEED sampler running in about 14.55 seconds against a baseline of 26.51 seconds with the standard euler sampler on the same setup, so roughly 1.82 times faster in that example. To use it, you drop the folder into ComfyUI's custom_nodes directory, restart ComfyUI, and then connect the new Sampler SPEED node into a workflow in the same way as any other sampler. The node exposes several inputs. base_sampler picks the underlying solver, with options like euler, heun, dpmpp_2m, and others. start_scale and mid_scale set the two intermediate resolutions, for example 0.5 and 0.75 of the final size. transition_1 and transition_2 control when the generation jumps to those bigger sizes. A taper value softens the seam between resolution steps to reduce ringing. The author warns that visible artifacts can appear depending on model and prompt, and that combining the node with torch.compile actually made things slower in their tests. Credit is given to the original paper authors and to a contributor who added the pluggable sampler dispatch and the cosine taper. The repository has 20 stars and is written in Python.