T

asagi4 e22008619b Linear adaptive guidance experiment

I think the implementation mostly matches the paper, but without proper betas
for the linear estimator, this doesn't actually do anything useful, and my
math-fu isn't good enough to calculate them.

2024-04-17 20:34:12 +03:00

__init__.py

Linear adaptive guidance experiment

2024-04-17 20:34:12 +03:00

.gitignore

Initial commit

2024-04-16 23:05:12 +03:00

LICENSE

add

2024-04-16 23:06:30 +03:00

README.md

Initial commit

2024-04-16 23:05:12 +03:00

README.md

Adaptive Guidance for ComfyUI

An implementation of adaptive guidance for ComfyUI

See https://bcv-uniandes.github.io/adaptiveguidance-wp/

What

There's an AdaptiveGuidance node (under sampling/custom_sampling/guiders) that can be used with SamplerCustomAdvanced. Normally, you should keep the threshold quite high, between 0.99 and 1.0

The node calculates the cosine similarity between the u-net's conditional and unconditional output ("positive" and "negative" prompts) and once the similarity crosses the specified threshold, it sets CFG to 1.0, effectively skipping negative prompt calculations and speeding up inference.

I'm not sure if the cosine similarity calculation matches the original paper since I had to translate from maths to Python, but it appears to work.