small.yaml

# Dimensions of the final video
width: 512
height: 512
# Frame rate of the final video
frames_per_second: 18

# When set, the runner will only simulate running all frames and output the configured additional context for inspecting ahead of actually running the scenario or for debugging.
# simulate_output: G:\shit\latent\out\vanishing.csv

output_path: outputs/small/
# A directory for intermediary results, and misc. other models that are required
persistence_dir: persistence/

# Torch device to use
torch_device: cuda
scenes:
  - # The prompt to generate images from
    # This example shows the prompt modulation/templating feature for modulating prompts during a scene to increase variety and allow for gradual changes throughout a scene
    # {{round(1-(scene_progress**0.9), 2)}} is evaluated at run-time, the purpose in this example is to have a slope on the "owl" and "cat" parts of the prompt,
    # modulating their attention strength throughout the scene using auto's attention syntax https://github.com/AUTOMATIC1111/stable-diffusion-webui-feature-showcase#attention
    # the template syntax used here is jinja2. some additional basic examples: https://jinja.palletsprojects.com/en/3.1.x/nativetypes/#examples
    prompt: a beautiful claymation artwork of an (owl:{{round(1-(scene_progress**0.9), 2)}}) (cat:{{round((scene_progress**0.9), 2)}}) in a beautiful claymation scene
    # How long this prompt should last
    duration: 3s
    # Interpolation between the previous and this scene. does not add to the duration of the scene.
    # NOTE: there is a bug with float durations here. Just use 1500ms instead of 1.5s instead as mitigation. https://github.com/sbaier1/pyttv/issues/16
    interpolation: 0s
    # Which mechanism to use to generate frames. Must be defined with parameters and type below
    mechanism: api
    # Optionally, some parameters of a mechanism may be modified at the scene-level. For example the number of steps to sample for in a diffusion model.
    mechanism_parameters:
      # Override the strength_schedule for this scene specifically.
      strength_schedule: min(1.0, 0.39+(is_turbo_step*0.12))
      # CFG scale
      scale: 10.5


mechanisms:
  - # Name of this parameter set within the config. Can be referenced in scenes
    name: api
    # Name of the mechanism to use
    type: api
    # Arbitrary parameters the text2video mechanism can take. These are always implementation specific
    mechanism_parameters:
      host: http://localhost:7860
      seed: 3455296820
      # How many turbo steps (lower amount of sampling steps) to run in-between "full sized" frames
      turbo_steps: 3
      # How many sampling steps to run for a turbo frame
      turbo_sampling_steps: 10
      # Noise schedule (how much noise to add to in-between frames), can be a function or constant
      noise_schedule: 0.02

      steps: 15
      # Euler a, LMS, ...
      sampler: Euler a
      # CFG scale
      scale: 9.5
      # This is denoising strength here, lower = better flow
      strength_schedule: (min(1.0, 0.76+(is_turbo_step*0.12))) #(sin(6*t)**20)*0.2
      # Type of animation to apply between frames, if any
      animation: 3D
      # Parameters for the animation
      animation_parameters:
        translation_x: 0.2
        translation_y: 0.4
        translation_z: 0.5
        # depth model parameters
        near_plane: 200
        far_plane: 10000
        fov: 40
        sampling_mode: bicubic
        padding_mode: border


# Other multi-modal context
additional_context:
  # This example uses no additional inputs
  input_mechanisms: []