diffusers
diffusers copied to clipboard
fix min-snr implementation
What does this PR do?
Fixes the implementation of min-snr training for v-prediction models
Based on implementation seen here. https://github.com/kohya-ss/sd-scripts/blob/main/library/custom_train_functions.py#L66
These two graphs show the weight scheme based on timestep from the kohya implementation.
the current implementation in diffusers takes the eps weight and increases everything by + 1, it appears that the +1 should be relocated to the denominator
- Schedulers: @yiyixuxu
- Training examples: @sayakpaul