DiffusionVideoSuperResolutionBase<T>
Base Classes · AiDotNet.Video
Base class for DIFFUSION-based video super-resolution models.
Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → VideoDiffusionModelBase<T> → DiffusionVideoSuperResolutionBase<T>
Implements: IVideoSuperResolution<T>
For Beginners
some video upscalers work by repeatedly “denoising” a noisy picture until a sharp frame emerges (diffusion), rather than transforming pixels in a single pass. Those models need two things at once: the diffusion machinery (a latent space and a denoising loop) and the super-resolution contract (an upscale factor). This base provides both.
How It Works
Why this class exists. Six models in the library are diffusion-based video super-resolution —
Stream-DiffVSR, StableVideoSR, MGLD-VSR, DOVE, SeedVR and FlashVSR — yet all of them derived from
VideoSuperResolutionBase, which has no latent space, no noise scheduler and no
denoiser. Each therefore either reimplemented diffusion privately or, more commonly, silently did
not implement it at all: Stream-DiffVSR was a plain conv/resblock/upsample CNN trained with one
feed-forward MSE pass, with none of its paper’s 4-step distilled denoiser, rollout distillation,
ARTG or TPM decoder.
C# allows a single base class, so the super-resolution capability is expressed as
IVideoSuperResolution — implemented by both this class and
VideoSuperResolutionBase — while the diffusion machinery is inherited from
VideoDiffusionModelBase. Consumers and the test-scaffold generator can then detect
the capability by interface rather than by matching a base-class NAME, which silently excludes any
model whose base is named differently.
Derived models must supply PredictNoise — the denoiser — because
that is what makes a diffusion model a diffusion model. A model that cannot provide one does not
belong on this base.
Properties
| Property | Summary |
|---|---|
int ScaleFactor { get; set; } |
Methods
| Method | Summary |
|---|---|
Tensor<T> EstimateFlow(Tensor<T> frame1, Tensor<T> frame2) |
|
Tensor<T> Upscale(Tensor<T> lowResFrames) |