Pre Training

I gathered 40 HD images of Margot Robbie. I am only using Birme to crop the HD photos instead of using Faceswap to align the faces. Some of the images were full body as I wanted to retain her face even when zoomed out. I also flipped the images horizontally to increase the number of images to train on. This should increase variety. These are 512x512 images instead of 1024x1024 images because I don’t have the specs to train a 1024x1024 model. I used Blip captioning to generate the filewords and edited each individually to reduce potential hallucinations.

Training

I used 0.005:100:0.0025:250,0.001:500,0.0005:1000,0.00025 for my learning rate. I am going for 10K training steps total. I am using a batch size of 1 with Gradient Accumulation Steps set to 3. I am paying $0.465/hour for 1xRTX 4090. I am using 11.2 out of 24 GB of VRAM. The estimated time of completion is 2 hours. For the embedding I am using 4 vectors per token. I switched to SD 1.5 EMA Only model for training.

Things that I could have done better

I could have upscaled the images before extracting the faces so I could reduce

blur.