/freeu_sft

An experiment, backpropogating FreeU loss back upstream.

Primary LanguageJupyter NotebookMIT LicenseMIT

This is a proof of concept, creating a stable diffusion pipeline that backpropogates the error compared to a FreeU (https://github.com/ChenyangSi/FreeU) inference.

The initial results are positive. My theory is that it can denoise our latent space.