MoG<T>
Models & Types · AiDotNet.Video.FrameInterpolation
MoG: motion-aware generative frame interpolation combining flow and diffusion.
Inheritance: NeuralNetworkBase<T> → VideoNeuralNetworkBase<T> → FrameInterpolationBase<T> → MoG<T>
For Beginners
MoG combines two approaches: first it figures out how things move (optical flow), then uses a generative AI model (diffusion) to “paint” the intermediate frame guided by that motion information. This produces sharper results than just blending warped frames, especially for complex motions.
Usage:
How It Works
MoG (2025) combines flow estimation with diffusion-based generation:
- Motion-aware conditioning: first estimates bidirectional optical flow using an EMA-VFI-style
flow network, then uses the estimated flows as spatial conditioning for a diffusion model rather than directly warping frames
- Flow-conditioned diffusion: the denoising U-Net receives concatenated flow maps as
additional input channels, guiding the diffusion process to generate motion-consistent intermediate frames with fine texture details
- Generative refinement: instead of blending warped frames (which can produce ghosting),
the diffusion model generates the intermediate frame from scratch, conditioned on the input frames and estimated motion, producing sharp results even in occluded regions
- Progressive denoising: multi-step denoising with motion-aware noise scheduling that
preserves motion coherence in early steps and refines textures in later steps
Reference: “MoG: Motion-Aware Generative Frame Interpolation” (2025)
Constructors
| Constructor | Summary |
|---|---|
MoG(NeuralNetworkArchitecture<T> architecture, MoGOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) |
Creates a MoG model in native training mode. |
MoG(NeuralNetworkArchitecture<T> architecture, string modelPath, MoGOptions options) |
Creates a MoG model in native training mode. |
Methods
| Method | Summary |
|---|---|
ModelMetadata<T> GetModelMetadata() |
|
Tensor<T> Interpolate(Tensor<T> frame0, Tensor<T> frame1, double t) |