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:
- Generating a base noise sequence for the window size
- Shifting and blending noise for extended frames
- 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. |