Overview
Generates temporally-correlated noise for the second pass of the VOID video refinement process. It takes the output video from Pass 1 and warps Gaussian noise along optical flow vectors, creating noise that moves consistently with the video content. This warped noise is used as the starting latent for Pass 2, which improves temporal consistency in the final output.Inputs
| Parameter | Description | Data Type | Required | Range |
|---|---|---|---|---|
optical_flow | Optical flow model from OpticalFlowLoader (RAFT-large). | OPTICAL_FLOW | Yes | - |
video | Pass 1 output video frames [T, H, W, 3]. | IMAGE | Yes | - |
width | Width of the output latent (default: 672). | INT | Yes | 16 to MAX_RESOLUTION (step 8) |
height | Height of the output latent (default: 384). | INT | Yes | 16 to MAX_RESOLUTION (step 8) |
length | Number of pixel frames. Rounded down to make latent_t even (patch_size_t=2 requirement), e.g. 49 → 45 (default: 45). | INT | Yes | 1 to MAX_RESOLUTION (step 1) |
batch_size | Number of identical noise sequences to generate (default: 1). | INT | Yes | 1 to 64 |
length parameter: The length value is automatically rounded down to the nearest valid value that produces an even latent_t dimension. This is required by the CogVideoX-Fun-V1.5 model’s patch_size_t=2 constraint. A warning is logged when rounding occurs.
Outputs
| Output Name | Description | Data Type |
|---|---|---|
warped_noise | A 5D tensor (B, C, T, H, W) containing optical-flow warped Gaussian noise, ready for use as the initial latent in VOID Pass 2. | LATENT |
This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub
Source fingerprint (SHA-256):
fa7f08bd5e050324d1d9681243e82d66b1613b8694a5b69a5949b39974c6bddb