We needed several highly stylized fashion photos for a tech demo about custom styles with Stable Diffusion. The images needed to look like photos while maintaining a slight surrealism.
All of the standard generations we tried weren't giving us the look we wanted, so the only option was to train a new custom model.
The Solution
We started by generating several colorful desert landscapes using a fine-tuned SDXL model in ComfyUI. Once we had about 50 that we liked, we used that as the dataset for training the model. The goal was to slightly overfit the model in order to get a bit of surrealism in the finished product.
Once we had our custom model trained, we simply added things to the prompt such as "surreal fashion photography" or "a dreamscape fantasy fashion photoshoot". We tried a variety of cameras and lenses in our prompts as well until we landed on the exact aesthetic we were going for.
