Skip to content

RF Diffusion Denoiser Config

Builds a denoiser configuration for RF Diffusion, controlling how much noise is applied to coordinates and frames during the diffusion steps. Lets you set initial and final noise scales and choose the interpolation schedule for both translation (Cα) and rotation (frame) noise.
Preview

Usage

Use this node to create a denoiser_config JSON and connect it to the 'RF Diffusion' node's denoiser_config input. Tune the noise scales to balance exploration (higher noise) vs. stability (lower noise), and pick a schedule (constant or linear) for how noise changes from the first to the final step.

Inputs

FieldRequiredTypeDescriptionExample
noise_scale_caTrueFLOATAmount of translation (Cα) noise at the initial step (t=T). 0.0–1.0 where higher values add more noise.1.0
final_noise_scale_caTrueFLOATAmount of translation (Cα) noise at the final step (t=1). 0.0–1.0 where higher values add more noise.1.0
ca_noise_schedule_typeTrueSTRINGInterpolation schedule for translation (Cα) noise between initial and final scales. Allowed values: 'linear', 'constant'.constant
noise_scale_frameTrueFLOATAmount of rotation (frame) noise at the initial step (t=T). 0.0–1.0 where higher values add more noise.1.0
final_noise_scale_frameTrueFLOATAmount of rotation (frame) noise at the final step (t=1). 0.0–1.0 where higher values add more noise.1.0
frame_noise_schedule_typeTrueSTRINGInterpolation schedule for rotation (frame) noise between initial and final scales. Allowed values: 'linear', 'constant'.constant

Outputs

FieldTypeDescriptionExample
denoiser_configJSONA configuration dictionary for the denoiser component, to be fed into the RF Diffusion node.{'noise_scale_ca': 1.0, 'final_noise_scale_ca': 0.8, 'ca_noise_schedule_type': 'linear', 'noise_scale_frame': 1.0, 'final_noise_scale_frame': 0.8, 'frame_noise_schedule_type': 'linear'}

Important Notes

  • Value ranges: All noise scale fields must be between 0.0 and 1.0.
  • Schedules: 'constant' keeps the same noise throughout; 'linear' interpolates from initial to final values across steps.
  • Integration: This node only affects generation if its JSON output is connected to the RF Diffusion node's denoiser_config input.
  • Tuning guidance: Lowering final noise scales can stabilize late steps; too-low values may reduce diversity or cause suboptimal designs.

Troubleshooting

  • Config not applied: Ensure the denoiser_config output is connected to the RF Diffusion node and that the run uses this connection.
  • Invalid value errors: Keep all noise scales within 0.0–1.0 and schedules to 'linear' or 'constant'.
  • Results too noisy or unstable: Decrease noise_scale_ and/or use 'linear' schedules with lower final_ values.
  • Results lack diversity: Increase initial noise_scale_* or use 'constant' schedules to maintain higher noise throughout.