paralleldomain.data_lab.config.reactor

class ReactorConfig(reactor_object, number_of_inference_steps=20, guidance_scale=7.5, noise_scheduler='dpm++', use_color_matching=False, use_camera_noise=False, clip_score_threshold=18, cameras_to_use=<factory>, random_seed=42, negative_prompt='comic style, drawing, motion blur, gray cube, white sign, warped, artifacts', inference_resolution=512, context_scale=2.0, color_matching_strength=0.5, undistort_input=True, undistort_context_scale_pad_factor=1.5, noise_strength=0.5, distortion_lookups=None)

Parameters for Reactor.

Parameters:

  • reactor_object (ReactorObject) –

  • number_of_inference_steps (int) –

  • guidance_scale (float) –

  • noise_scheduler (str) –

  • use_color_matching (bool) –

  • use_camera_noise (bool) –

  • clip_score_threshold (float) –

  • cameras_to_use (List) –

  • random_seed (int) –

  • negative_prompt (str) –

  • inference_resolution (int) –

  • context_scale (float) –

  • color_matching_strength (float) –

  • undistort_input (bool) –

  • undistort_context_scale_pad_factor (float) –

  • noise_strength (float | None) –

  • distortion_lookups (Dict[str, DistortionLookupTable] | None) –

reactor_object

defines the object to be modified using a prompt

number_of_inference_steps

diffusion steps of the image generation model

guidance_scale

controls how much the image generation process follows the text prompt

noise_scheduler

noise scheduler of the image generation model

use_color_matching

applies hue color transform on the generated object

use_camera_noise

applies camera noise on the generated object

clip_score_threshold

rejects generated samples with a clip score below threshold. Good range between 18-26.

cameras_to_use

which cameras to use for Reactor

random_seed

is passed to the Reactor model

negative_prompt

helps guide the Reactor diffusion process.

inference_resolution

image crop will be scaled to this resolution before the image generation process

context_scale

amount of context to add around the object instance mask.

context_scale = 1 no additional context.
context_scale 2 = add padding such that resulting crop is twice as large as the original mask.
color_matching_strength

strength of color matching. 0.0 = no strength, 1.0 = full strength.

undistort_input

whether to undistort fisheye inputs

undistort_context_scale_pad_factor

how much padding to add before undistorting

noise_strength

strength of noise applied to the image generation model, only used when change_shape = False

distortion_lookups

maps cameras_to_use to DistortionLookupTable. Required for sensor rigs with pd_fisheye cameras. Helper function: extract_distortion_lookup_tables

class ReactorObject(prompts, new_class_id=None, asset_name=None, change_shape=True, replacement_probability=1.0, registry_class_name=None, registry_class_id=None)
Parameters:

  • prompts (List[str]) –

  • new_class_id (int) –

  • asset_name (str | None) –

  • change_shape (bool) –

  • replacement_probability (float) –

  • registry_class_name (str | None) –

  • registry_class_id (int) –

prompts

prompts are guiding the Reactor process towards which objects to generate

new_class_id

class_id to assign to generated object

asset_name

target asset name that provides object mask for Reactor

change_shape

True: mode to convert primitive shapes (e.g. a cube) into a new object with different texture and shape. False: mode to texturize an existing object, while maintaining the same shape.

replacement_probability

When set to 1.0 all object instances will be replaced. only used when change_shape = False

registry_class_name

object instances of this class name will be texturized by Reactor.

registry_class_id

class_id stored in the asset registry, will be retrieved automatically