FreeNoiseModule<T>

Models & Types · AiDotNet.Diffusion.Acceleration

FreeNoise module for tuning-free longer video generation via noise rescheduling.

For Beginners

The FreeNoise module rearranges noise patterns to enable longer video generation from models trained on short clips. It is a simple, training-free technique that extends video length without quality loss.

How It Works

References:

  • Paper: “FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling” (Qiu et al., 2024)

FreeNoise enables generating longer videos from short-video diffusion models without any fine-tuning. The key idea is noise rescheduling: instead of using independent random noise for all frames, FreeNoise constructs temporally correlated noise by:

  1. Generating a base noise sequence for the window size
  2. Shifting and blending noise for extended frames
  3. Using window-based attention with shared noise patterns

This maintains temporal consistency across windows while extending generation length.

Constructors

Constructor Summary
FreeNoiseModule(int windowSize, int noiseShiftStride, double blendRatio, int? seed) Initializes a new FreeNoise module.

Properties

Property Summary
double BlendRatio { get; } Gets the blend ratio between shifted and fresh noise.
int NoiseShiftStride { get; } Gets the noise shift stride.
int WindowSize { get; } Gets the window size for noise generation.

Methods

Method Summary
Tensor<T> GenerateRescheduledNoise(int targetFrames, int[] latentShape) Generates temporally correlated noise for a target number of frames.
Tensor<T> GetBaseNoise() Gets a copy of the stored base noise from the last generation.
void Reset() Resets the module state.