Sign in

Nodes Browser

ComfyDeploy: How ComfyUI-RegionalAdaptiveSampling works in ComfyUI?

What is ComfyUI-RegionalAdaptiveSampling?

ComfyUI implementation of Regional Adaptive Sampling, (original implementation at https://github.com/microsoft/RAS).

How to install it in ComfyDeploy?

Head over to the machine page

  1. Click on the "Create a new machine" button
  2. Select the Edit build steps
  3. Add a new step -> Custom Node
  4. Search for ComfyUI-RegionalAdaptiveSampling and select it
  5. Close the build step dialig and then click on the "Save" button to rebuild the machine

Regional Adaptive Sampling

Regional Adaptive Sampling is a new technique for accelerating the inference of diffusion transformers. It essentially works as a KV Cache inside the model, picking regions that are likely to be updated by each diffusion step and passing in only those tokens.

This implementation is simple to use, and compatible with Flux (dev & schnell) at HunYuanVideo. I may add support for other models in the future.

Usage

Apply the Regional Adaptive Sampling node to the desired model. It has the following parameters:

  • sample_ratio: The percent of tokens to keep in the model on a RAS pass. Anything below 0.3 is usually very bad quality.
  • warmup_steps: The number of steps to do without RAS at the beginning. Setting higher will decrease the speedup, and setting it lower will degrade the composition.
  • hydrate_every: Every hydrate_every steps, we do a full run through the model with all tokens, to refresh the stale cache. Set to 0 to disable and do full RAS.
  • starvation_scale: Controls how the model decides which part of the image to focus on. Increasing it will probably shift quality from the main subject to the background. The default of 0.1 is what's used in the paper, and I haven't tried anything else.