| name: stable-diffusion-train | |
| infer: | |
| unconditional_guidance_scale: 3 | |
| num_images_per_prompt: 4 | |
| hint_image_size: 512 | |
| height: 512 | |
| width: 512 | |
| down_factor: 8 | |
| inference_steps: 50 | |
| sampler_type: 'DDIM' | |
| eta: 0 | |
| output_type: 'pil' | |
| save_to_file: True | |
| out_path: 'controlnet' | |
| seed: 355 | |
| prompts: | |
| - high quality picture of a house in oil painting style | |
| control: | |
| - /datasets/coco-stuff/house.png #images/val2017/000000001584.jpg | |
| # Depending on the input control, if the input control is already the conditioning image, null should be passed here | |
| # If a reconstruction target is used as control, then preprocessing function that turns it into a conditioning image needs to be specified | |
| control_image_preprocess: | |
| trainer: | |
| devices: 1 | |
| num_nodes: 1 | |
| accelerator: gpu | |
| precision: 16 | |
| logger: False # logger provided by exp_manager | |
| model: | |
| restore_from_path: /ckpts/controlnet/30k.nemo | |
| precision: ${trainer.precision} | |
| strength: 2.0 | |
| guess_mode: False |