FateZero/config/low_resource_teaser/jeep_watercolor_ddim_10_steps_disk_store.yaml

# CUDA_VISIBLE_DEVICES=0 python test_fatezero.py --config config/low_resource_teaser/jeep_watercolor_ddim_10_steps_disk_store.yaml

pretrained_model_path: "./ckpt/stable-diffusion-v1-4"

train_dataset:
    path: "data/teaser_car-turn"
    prompt: "a silver jeep driving down a curvy road in the countryside"
    n_sample_frame: 8
    sampling_rate: 1
    stride: 80
    offset: 
        left: 0
        right: 0
        top: 0
        bottom: 0


validation_sample_logger_config:
    use_train_latents: true
    use_inversion_attention: true
    guidance_scale: 7.5
    source_prompt: "${train_dataset.prompt}"
    prompts: [
        # a silver jeep driving down a curvy road in the countryside,
        watercolor painting of a silver jeep driving down a curvy road in the countryside,
    ]
    p2p_config:
        0: 
            # Whether to directly copy the cross attention from source 
            # True: directly copy, better for object replacement
            # False: keep source attention, better for style
            
            is_replace_controller: False

            # Semantic layout preserving. High steps, replace more cross attention to preserve semantic layout
            cross_replace_steps: 
                default_: 0.8

            # Source background structure preserving, in [0, 1]. 
            # e.g., =0.6 Replace the first 60% steps self-attention
            self_replace_steps: 0.8

            
            # Amplify the target-words cross attention, larger value, more close to target
            eq_params: 
                words: ["watercolor"]
                values: [10,10] 
            
            # Target structure-divergence hyperparames
            # If you change the shape of object better to use all three line, otherwise, no need.
            # Without following three lines, all self-attention will be replaced
            # blend_words: [['jeep',], ["car",]] 
            # masked_self_attention:  True
            # masked_latents: False   # performance not so good in our case, need debug            
            # bend_th: [2, 2]
            # preserve source structure of blend_words , [0, 1]
            # default is bend_th: [2, 2]  # replace full-resolution edit source with self-attention 
            # bend_th-> [0.0, 0.0], mask -> 1, use more edit self-attention, more generated shape, less source acttention
            

    clip_length: "${..train_dataset.n_sample_frame}"
    sample_seeds: [0]

    num_inference_steps: 10
    prompt2prompt_edit: True

disk_store: True
model_config:
    lora: 160
    # temporal_downsample_time: 4
    SparseCausalAttention_index: ['mid'] 
    least_sc_channel: 640
    # least_sc_channel: 100000

test_pipeline_config:
    target: video_diffusion.pipelines.p2pDDIMSpatioTemporalPipeline.p2pDDIMSpatioTemporalPipeline
    num_inference_steps: "${..validation_sample_logger.num_inference_steps}"

epsilon: 1e-5
train_steps: 10
seed: 0
learning_rate: 1e-5
train_temporal_conv: False
guidance_scale: "${validation_sample_logger_config.guidance_scale}"