AiModelResultRadianceFieldExtensions

Helpers & Utilities · AiDotNet.NeuralRadianceFields.Extensions

Model-specific inference extensions on AiModelResult for the radiance-field family (NeRF, InstantNGP, GaussianSplatting). Fixes #1836.

How It Works

Model is internal by design (IP protection — same reason model save/load is license-gated). Consumers outside the AiDotNet assembly can’t retrieve the trained model from BuildAsync’s return, so previously the only way to call radiance-field-specific inference (RenderImage, RenderRays, QueryField) was to keep a raw reference to the model passed to ConfigureModel — fragile under factories, callbacks, or DeepCopy.

These extension methods live in the same assembly as AiModelResult so they access the internal Model property directly. Callers just add using AiDotNet.NeuralRadianceFields.Extensions; and the family-specific inference methods appear on the result.

Beyond industry standard. Reference facades (PyTorch skorch, Keras wrappers) either force everything through a generic Predict OR expose the raw model publicly. Neither is great: generic Predict loses model-specific inference (rendering a camera view, sampling from a diffusion model, generating text) and public model exposure defeats IP protection. Namespaced extensions give the best of both: hidden model + discoverable domain-specific inference.

Methods

Method Summary
Task<Tensor<T>> RenderImageAsync(AiModelResult<T, TInput, TOutput> result, Vector<T> cameraPosition, Matrix<T> cameraRotation, int imageWidth, int imageHeight, T focalLength, CancellationToken cancellationToken) Async variant of RenderImage for orbit-trajectory renders where the caller wants to render many frames in parallel without blocking.
Tensor<T> RenderImage(AiModelResult<T, TInput, TOutput> result, Vector<T> cameraPosition, Matrix<T> cameraRotation, int imageWidth, int imageHeight, T focalLength) Renders a full RGB image by casting rays from the camera and volume-rendering through the radiance field.
Tensor<T> RenderRays(AiModelResult<T, TInput, TOutput> result, Tensor<T> rayOrigins, Tensor<T> rayDirections, int numSamples, T nearBound, T farBound) Renders rays through the radiance field using volume rendering.
Tensor<T>[] RenderImageBatch(AiModelResult<T, TInput, TOutput> result, IEnumerable<ValueTuple<Vector<T>, Matrix<T>, int, int, T>> views) Batched variant that renders N views in parallel using the .NET thread pool.
ValueTuple<Tensor<T>, Tensor<T>> QueryField(AiModelResult<T, TInput, TOutput> result, Tensor<T> positions, Tensor<T> viewingDirections) Queries the raw radiance field at 3D positions (returns per-point RGB + density without volume rendering).