There are detailed changes below the next paragraph, you might need to expand this version changes box!
I'm uploading this "WIP version" of v7 as I'm not sure when / if I'll have the time / motivation to properly finish it up. It is roughly equivalent to v6 with the changes to the training process from v7 of my Dawn LoRA. It already fixes quite a few issues with v6 and should usually perform better. But it is not at the same level of improvement yet. If I'm not doing a "proper" v7 with major dataset changes, I might simply rename this version in the future.
Most of the changes (look at LoRA metadata for more details):
Updated to much newer Kohya scripts version (>3 months newer)
Updated training images
Switched to keep original-ish aspect ratios (slightly cropped and resized to compatible SD 1.5 resolutions) where it makes sense instead of forcing squares
Added regularization images to actually make trigger tag work correctly
1 regularization image per training image
Generated by the base model at the same resolution, with the same tags (minus the activation tag)
Currently at half loss during training as they had a bit too much influence otherwise
With this, the LoRA now reverts back much closer to base model knowledge without the activation tag (which is correct!)
Might care a bit more about the positioning of the activation tag now as it was trained with "keep tokens" to keep the activation token at the front when shuffling
Normalized repeats to 1 (only using different values if ever in need of balancing datasets) and learning rates back to defaults
Compensated with different epoch settings
Reverted back to training at 128 dims and a resize down to 32
Results were better across the board and the resizing also removes a bit of noise as a bonus
Used NO dynamic resize method as results for sv_fro@0.99 and sv_ratio@20 did not seem too different from or better than a simple resizing
Added training warmup of 10%
Not sure if this had much impact, might remove again in the future
Added "scale weight norms" with value 1 during training
Supposedly helps against overfitting and might make LoRAs more compatible with others
Used FreeU extension for example image generation to further improve results
There are detailed changes below the next paragraph, you might need to expand this version changes box!
I'm uploading this "WIP version" of v7 as I'm not sure when / if I'll have the time / motivation to properly finish it up. It is roughly equivalent to v6 with the changes to the training process from v7 of my Dawn LoRA. It already fixes quite a few issues with v6 and should usually perform better. But it is not at the same level of improvement yet. If I'm not doing a "proper" v7 with major dataset changes, I might simply rename this version in the future.
Most of the changes (look at LoRA metadata for more details):
Updated to much newer Kohya scripts version (>3 months newer)
Updated training images
Switched to keep original-ish aspect ratios (slightly cropped and resized to compatible SD 1.5 resolutions) where it makes sense instead of forcing squares
Added regularization images to actually make trigger tag work correctly
1 regularization image per training image
Generated by the base model at the same resolution, with the same tags (minus the activation tag)
Currently at half loss during training as they had a bit too much influence otherwise
With this, the LoRA now reverts back much closer to base model knowledge without the activation tag (which is correct!)
Might care a bit more about the positioning of the activation tag now as it was trained with "keep tokens" to keep the activation token at the front when shuffling
Normalized repeats to 1 (only using different values if ever in need of balancing datasets) and learning rates back to defaults
Compensated with different epoch settings
Reverted back to training at 128 dims and a resize down to 32
Results were better across the board and the resizing also removes a bit of noise as a bonus
Used NO dynamic resize method as results for sv_fro@0.99 and sv_ratio@20 did not seem too different from or better than a simple resizing
Added training warmup of 10%
Not sure if this had much impact, might remove again in the future
Added "scale weight norms" with value 1 during training
Supposedly helps against overfitting and might make LoRAs more compatible with others
Used FreeU extension for example image generation to further improve results
Recommended weight: 1
Training model: Anything V3
Example image generation model: AbyssOrangeMix2 - Hardcore