A test version of latex kigurumi Lora, trained on a 3090Ti with about 18k pics scraped from Twitter, then tagged with Mistral's Pixtral-Large-2411

Current Lora config: 128 Rank Linear + 32 Rank Conv for SDXL version

Trained at 1024px, 64 batch effective (8 batch x 8 accumulation), Prodigy with cosine annealing.

Upcoming version:

LyCORIS GLoRA + DoRA, 1024px + 1536px multires