FateZero/config/teaser/jeep_posche.yaml

# CUDA_VISIBLE_DEVICES=0 python test_fatezero.py --config config/teaser/jeep_posche.yaml

pretrained_model_path: "./ckpt/jeep_tuned_200"

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
    prompts: [
        a silver jeep driving down a curvy road in the countryside,
        a Porsche car 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: 0.9

            
            # Amplify the target-words cross attention, larger value, more close to target
            # Usefull in style editing
            eq_params: 
                words: ["watercolor", "painting"]
                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
            # Usefull in shape editing
            blend_words: [['jeep',], ["car",]] 
            masked_self_attention:  True
            # masked_latents: False   #  Directly copy the latents, performance not so good in our case           
            
            # preserve source structure of blend_words , [0, 1]
            # bend_th-> [1.0, 1.0], mask -> 0, use inversion-time attention, the structure is similar to the input
            # bend_th-> [0.0, 0.0], mask -> 1, use more edit self-attention, more generated shape, less source acttention
            bend_th: [0.3, 0.3]

        1:
            cross_replace_steps: 
                default_: 0.5
            self_replace_steps: 0.5
            
            use_inversion_attention: True
            is_replace_controller: True

            blend_words: [['silver', 'jeep'], ["Porsche", 'car']] # for local edit. If it is not local yet - use only the source object: blend_word = ((('cat',), ("cat",))).
            masked_self_attention: True
            bend_th: [0.3, 0.3]
            
    clip_length: "${..train_dataset.n_sample_frame}"
    sample_seeds: [0]

    num_inference_steps: 50
    prompt2prompt_edit: True

    
model_config:
    lora: 160


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}"