NOTE: V4 changed the quality tags. Update your prompts.
This LoRA adds quality tags inspired by NAI to create aesthetically pleasing images without artist tags. The dataset was split into 4 tag groups based on their favcount on e621.
V4 is for fluffyrock-1088-minsnr-zsnr-vpred-ema (trained on e267), which requires a corresponding yaml file. You will also need the CFG Rescale extension.
V3 is for fluffyrock vpred (trained on e160).
Usage: add the tags to the prompt, use emphasis to change the strength of the effect. Do NOT use LoRA weight.
Positive prompt: (best quality, good quality:1.2)
Negative prompt: (worst quality, bad quality:1.2)Description
Trained for 3 more epochs. v0.5 was mostly converged already, so the differences are minor.
FAQ
Comments (4)
Could you give a bit more detail on how this was trained? I've never seen a quality LoRA that works this well before.
The main thing was using the Prodigy optimizer with a constant learning rate schedule for most of the training, then switching to cosine schedule for the last ~25% to let it converge.
The score thresholds aren't evenly distributed either. They're percentiles corresponding to -2 stddev, -1, +1, +2. So masterpiece and worst quality are about 2% each, while normal quality is the middle ~67%
I also found that starting at a lower resolution (512px) learned the style quicker, so v2 was trained for 16 epochs (10k images per epoch) at 512px, then for another 8 epochs at 768px.
After more testing:
- The middle normal quality bucket isn't needed.
- The style developed much faster once I dropped artist and style-related meta tags.