name: test-run trainer: model_path: "nai_diffusers" # model_url: "" # init_batch_size: 12 batch_size: 6 resolution: 512 center_crop: false gradient_checkpointing: true clip_skip: 2 seed: 1138 use_ema: false use_hivemind: false use_xformers: true lr_scale: sqrt train_text_encoder: false attention_slicing: false vae_slicing: true offset_noise: true offset_noise_val: 0.06 min_snr: true min_snr_val: 3 checkpoint: monitor: 'epoch' dirpath: checkpoint_bench filename: 'sample-nd-epoch{epoch:02d}' auto_insert_metric_name: false every_n_epochs: 1 save_top_k: -1 mode: 'max' save_last: False lightning: accelerator: gpu devices: -1 auto_select_gpus: true # limit_train_batches: 100 max_epochs: 20 precision: 16 log_every_n_steps: 1 accumulate_grad_batches: 1 gradient_clip_val: 0 auto_scale_batch_size: false auto_lr_find: false arb: enabled: true debug: false base_res: [512, 512] max_size: [768, 512] divisible: 64 max_ar_error: 4 min_dim: 256 dim_limit: 1024 dataset: img_path: - "/home/ec2-user/SageMaker/minibench/images" center_crop: false ucg: 0.07 debug_arb: false num_workers: 6 process_tags: true important_tags: [] allow_duplicates: false augment: enabled: true debug: false dropout: 0.8 methods: adjust_brightness: 0.04 adjust_contrast: 0.05 rotate_with_crop: 3 optimizer: name: lion_pytorch.lion_pytorch.Lion params: lr: 1.7e-7 weight_decay: 1e-2 # eps: 1e-8 lr_scheduler: name: torch.optim.lr_scheduler.CosineAnnealingLR warmup: enabled: true init_lr: 2e-8 num_warmup: 50 strategy: "cos" params: T_max: 6 eta_min: 1e-7 last_epoch: -1 monitor: wandb_id: "" huggingface_repo: "" hf_auth_token: "" store_checkpoints: true sampling: enabled: true use_wandb: true every_n_steps: 112 every_n_epochs: 1 save_dir: "samples" seed: 1139 height: 768 width: 512 steps: 20 cfg_scale: 7 negative_prompts: - "nsfw, text, error, signature, watermark, username, realistic,3d, multiple people, animals, lowres, cropped, worth quality ,low quality, normal quality, jpeg artifacts, blurry, bad anatomy, bad hands, bad arms, bad feet, bad anatomy, missing fingers, extra digits, fewer digits, long neck, missing legs, huge person, optical_illusion" - "nsfw, text, error, signature, watermark, username, realistic,3d, multiple people, animals, lowres, cropped, worth quality ,low quality, normal quality, jpeg artifacts, blurry, bad anatomy, bad hands, bad arms, bad feet, bad anatomy, missing fingers, extra digits, fewer digits, long neck, missing legs, huge person, optical_illusion" - "nsfw, text, error, signature, watermark, username, realistic,3d, multiple people, animals, lowres, cropped, worth quality ,low quality, normal quality, jpeg artifacts, blurry, bad anatomy, bad hands, bad arms, bad feet, bad anatomy, missing fingers, extra digits, fewer digits, long neck, missing legs, huge person, optical_illusion" - "nsfw, text, error, signature, watermark, username, realistic,3d, multiple people, animals, lowres, cropped, worth quality ,low quality, normal quality, jpeg artifacts, blurry, bad anatomy, bad hands, bad arms, bad feet, bad anatomy, missing fingers, extra digits, fewer digits, long neck, missing legs, huge person, optical_illusion" - "nsfw, text, error, signature, watermark, username, realistic,3d, multiple people, animals, lowres, cropped, worth quality ,low quality, normal quality, jpeg artifacts, blurry, bad anatomy, bad hands, bad arms, bad feet, bad anatomy, missing fingers, extra digits, fewer digits, long neck, missing legs, huge person, optical_illusion" prompts: - "best quality, a girl with a yellow hat and a yellow shirt, sitting, white legwear, uniform, jacket, thighhighs, bare shoulders, high contrast, paint splatter" - "best quality, 1girl, solo, loli, cat girl, silver hair ,blue eyes, flat chest, solo, beautiful detailed background, messy hair, long hair" - "masterpiece, best quality, 1girl, a clean painting of a girl with long white hair drinking coffee in a coffee bar" - "best quality, 1girl, a clean painting of a girl with long white hair drinking coffee in a coffee bar" - "1girl, a clean painting of a girl with long white hair drinking coffee in a coffee bar" encoder: # Leave blank to load from model text: vae: