Type | |
Stats | 314 1,303 47 |
Reviews | (58) |
Published | Feb 25, 2024 |
Base Model | |
Training | Steps: 14,250 Epochs: 40 |
Usage Tips | Strength: 5 |
Trigger Words | A24 style Scene from a film |
Hash | AutoV2 7431E23438 |
This SDXL LoRA was trained on screenshots from six A24 Films productions including The Florida Project, Springbreakers, Moonlight, Midsommar, Under the Skin & Uncut Gems. The goal was to capture the overall look of A24 films which are known for their visual style, lighting, depth of field and anamorphic lens related bokeh. FFMPEG was used to extract one frame every four seconds from these films. From all frames I manually selected shots that had interesting lighting and composition. In total 285 screenshots were selected and captioned using CogVLM and the captions were edited down to remove verbose language. Training was for 40 epochs using prodigy optimizer with a batch size of 4 and a gradient accumulation of 2. Results are mixed. I suspect the disparate lighting styles of Midsommar and Springbreakers did not train well together. Despite that, interesting outputs are possible. I will follow up with separate LoRAs from the individual films.