This is a somewhat successful attempt to create an asari model for Stable Diffusion with Dreambooth. It is not generally useable. I leave my notes here in case it helps someone get one step closer to generating Mass Effect stuff.

asari3: Dreambooth with 53 samples

Tokens trained: asari, purple, blue, lightblue, green, teal, armor

The number starts on 3 because early experiments were met with absolute failure, which made me seek help on the SD discord where I got some great advice from Guizmus (link to HF page) and this is the result of putting that into practice.

The sample set is mostly in-game screenshots with minimal labeling. The games are not great at face, clothing and armor variety, which leads to overfitting facial features and not shaking off clothing/armor details in the token. No class/regularization images were used.

• asari3_2_8100: 150 epochs at LR 1E-6
• asari3_3_8100: 150 epochs at LR 5E-7
• asari3_3_13500: above + 100 epochs at LR 1E-7

Some possible improvements

• The ME: Legendary Edition games has a visual issue with asari eyes, which does show in the results. Maybe find a mod or graphics settings that reduce that and redo parts of the dataset.
• Label the images' clothes to mitigate the dataset's poor variety.
• Try to work in regularization again. It doesn't have too much bleed-over aside from some peculiar haircuts, so this might not be very helpful.
• Expand the dataset with realistic-style fanart.

asari3_4: Simplifying it

This was a training done with only the asari label on every image. This is the same 9 pictures generated with the models starting from the least trained (45 epochs) until the last increment of 45 that was possible before the overtraining artifacts set in.

There is a loss of color control: markings aren't easy to colorize by prompt, and multi-color faces are more common.

Notes

This is not related to the Asari Diffusion LoRa model on CivitAI.