<a rel="ugc" href="https://www.instagram.com/camilamorrone/?hl=en">Camila Morrone</a> is an American actress and model of Argentine descent.I gotta admit I'm a bit puzzled with this TI. After about 500 steps it was already overtrained, even though I think the dataset was carefully selected and edited... I blame the "Use PNG alpha channel as loss weight" option, which seems to accelerate the learning process by a lot, but... I'm not sure anymore. The thing is, even if a tad overcooked, the embedding seems really accurate and is producing really nice results for me in all of the models I've tried. If you get saturated images, try to reduce the embedding weight. Something like "(cm550:0.98)", which is a tiny decrease in the TI's weight, already makes a very noticeable difference in many cases. In any case, I hope you like this one! She's definitely a beauty, and I think base Stable Diffusion didn't do her justice.

Camila Morrone

Training breakdownDataset: 29 images, filewords (BLIP captions) edited manually4 tokens, initialization text: "*"Learning rate: 0.05:10, 0.02:20, 0.01:60, 0.005:200, 0.002:500, 0.001:1500, 0.0005Batch size: 3Gradient accumulation: 2Steps: 2000 (interrupted at about 1600, chose step 550 after checking several others)Sampling method: DeterministicUse PNG alpha channel as loss weight: ONShuffle tags by ',' when creating prompts: ONDrop out tags when creating prompts: 0.1

Files with SHA256: 59bc36c97a661a102482e0327f56874f5390da824e83e4deb8d8b034dc326d8e