The earlier versions of this model was my first SDXL LoRA attempts. This version is with my latest training methodology.
Dataset gathered and trained on a single concept (model had a short active career period).
~7.5 hour Kohya_ss training in 16530 steps across 19 epochs @base dim (rank): 32.
1248x1248px training resolution base
Used a wide variety of resolutions in training data, however by far the most common bucketed resolution was @ 1216x1216px.
Comparing results and my datasets since the v1 and v1.05 versions of this LoRA to my other LoRAs have given me even more ideas on quality improvements. Stay tuned...
The earlier versions of this model was my first SDXL LoRA attempts. This version is with my latest training methodology.
Dataset gathered and trained on a single concept (model had a short active career period).
~7.5 hour Kohya_ss training in 16530 steps across 19 epochs @base dim (rank): 32.
1248x1248px training resolution base
Used a wide variety of resolutions in training data, however by far the most common bucketed resolution was @ 1216x1216px.
Comparing results and my datasets since the v1 and v1.05 versions of this LoRA to my other LoRAs have given me even more ideas on quality improvements. Stay tuned...