LearnedPrior<T>
Base Classes · AiDotNet.NeuralRadianceFields.Data
Optional single-image / few-shot reconstruction prior attached to an ImageView.
#1834 excellence goal — reference impls (Zero123, TripoSR, SyncDreamer) all live in separate
codebases with their own inference stacks. Attaching a prior directly to the view unifies
single-image reconstruction into the same NeRF/GS training loop: the model composes the
prior’s hallucinated novel views with the caller’s actual photos.
Known derived types: IsotropicPrior<T>
How It Works
The default prior is null — normal multi-view reconstruction with no
hallucination. Users opt in by supplying a prior on the single-photo view. Callers can
implement ILearnedPrior<T> to plug in their own diffusion-based novel-view
synthesizers; the reference impl IsotropicPrior ships as a paper-standard
fallback that produces spatially-isotropic (view-independent) predictions.
Properties
| Property | Summary |
|---|---|
double Confidence { get; set; } |
Prior weight in [0, 1] — how strongly the model should trust this hallucination vs its own current estimate. |
Methods
| Method | Summary |
|---|---|
Tensor<T> SynthesizeNovelView(Vector<T> cameraPosition, Matrix<T> cameraRotation, int height, int width) |
Synthesizes a novel-view prediction for the given camera pose. |