jeep.yaml

# CUDA_VISIBLE_DEVICES=0 python ContextDiff_finetune.py --config config/jeep.yaml
pretrained_model_path: "ckpt/stable-diffusion-v1-5"

dataset_config:
    path: "data/jeep"
    prompt: "a silver jeep driving down a curvy road in the countryside,"
    n_sample_frame: 8
    # n_sample_frame: 22
    # class_data_root: "data/car"
    # class_data_prompt: "a photo of a car"

    sampling_rate: 1
    stride: 80
    offset: 
        left: 0
        right: 0
        top: 0
        bottom: 0

editing_config:
    use_invertion_latents: True
    use_inversion_attention: True
    guidance_scale: 7.5
    editing_prompts: [
        a silver jeep driving down a curvy road in the countryside,
        a Porsche car driving down a curvy road in the countryside,
        a silver jeep driving down a curvy road in the countryside covered with snow,
    ]
    clip_length: "${..dataset_config.n_sample_frame}"
    sample_seeds: [12734]
    
    num_inference_steps: 50 # 15 minutes
    strength: 0.99

trainer_pipeline_config:
    target: video_diffusion.trainer.ddpm_trainer.DDPMTrainer

test_pipeline_config:
    target: video_diffusion.pipelines.ddim_spatial_temporal.DDIMSpatioTemporalStableDiffusionPipeline

model_config:
    lora: 160
    # temporal_downsample_time: 4
    # SparseCausalAttention_index: [-1, 1, 'first', 'last'] 

enable_xformers: True
mixed_precision: 'fp16'
gradient_checkpointing: True

train_steps: 200
validation_steps: 50 
checkpointing_steps: 50
seed: 0
learning_rate: 1e-5
# prior_preservation: 1.0
train_temporal_conv: True