As mentioned in title, it seems like "adaptive_noise_scale = 0.2" in the toml file under training arguments gives an unrecognised argument error when running the the 5.5 Start training cell on Colab. Given this feature works in Kohya_ss and baltimis fork, I thought it would work without issue but seems not yet implemented? Would be super helpful to have this for larger loras and avoid overfit and artefacts.
As mentioned in title, it seems like "adaptive_noise_scale = 0.2" in the toml file under training arguments gives an unrecognised argument error when running the the 5.5 Start training cell on Colab. Given this feature works in Kohya_ss and baltimis fork, I thought it would work without issue but seems not yet implemented? Would be super helpful to have this for larger loras and avoid overfit and artefacts.