Even though the images used for training already contain an overweight character, I find the best way to get consistant body weight is to use chubby woman in the prompt. Here is an example of a prompt: <code>l34n0r, chubby woman, lying on a bed, wearing red silk pyjamas</code> First try on an overweight character LoRA created with z-image-turbo generated inputs. I find it hard to generate consistent body weight with z-image-turbo, and I wanted to see if with a LoRA trained on a dataset with an already chubby figure it would be easier (it kind of is). The dataset creation process: I first generated a prompt that would define the character exhaustively, locked the seed and generated about 200 images with various actions / lighting / clothes. I then do a pass of face recognition on the generated images to keep as much consistency as possible for the input images and select about 20 that are used for training.

Files with SHA256: 5b44b6c696601a3bd1f3ef2e8e8ff8a6ce4735bacbdb446a340b612a627a91d5