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


train_dataset:
    path: "data/style/red_water_lily_opening"
    prompt: "a pink water lily"
    start_sample_frame: 1
    n_sample_frame: 8
    # n_sample_frame: 22
    sampling_rate: 20
    stride: 8000
    # offset: 
    #     left: 300
    #     right: 0
    #     top: 0
    #     bottom: 0

validation_sample_logger_config:
    use_train_latents: True
    use_inversion_attention: True
    guidance_scale: 7.5
    prompts: [
        a pink water lily,
        Claude Monet painting of a pink water lily,
    ]
    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 preserving and replacement Debug me
            cross_replace_steps: 
                default_: 0.7

            # Source background structure preserving, in [0, 1]. 
            # e.g., =0.6 Replace the first 60% steps self-attention
            self_replace_steps: 0.7

            
            # Amplify the target-words cross attention, larger value, more close to target
            eq_params: 
                words: ["silver", "sculpture"]
                values: [2,2] 
            
            # 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: [['cat',], ["cat",]] 
            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]  # preserve all source self-attention 
            # bend_th : [0.0, 0.0], mask -> 1, use more att_replace, more generated attention, less source acttention
            

        1:
            is_replace_controller: False
            cross_replace_steps: 
                default_: 0.5
            self_replace_steps: 0.5
            
            eq_params: 
                words: ["Monet"]
                values: [10]
            
    clip_length: "${..train_dataset.n_sample_frame}"
    sample_seeds: [0]
    val_all_frames: False

    num_inference_steps: 50
    prompt2prompt_edit: True

    
model_config:
    lora: 160
    # temporal_downsample_time: 4
    SparseCausalAttention_index: ['mid'] 
    least_sc_channel: 1280
    # 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}"