GPT4Point<T>

Models & Types · AiDotNet.VisionLanguage.ThreeD

GPT4Point: unified point-language understanding and generation.

Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → GPT4Point<T>

Implements: IThreeDVisionLanguageModel<T>

For Beginners

GPT4Point is a vision-language model for understanding and generating 3D point clouds using natural language. Default values follow the original paper settings.

How It Works

GPT4Point (2024) is a unified framework for point cloud-language understanding and generation. It bridges 3D point cloud representations with language models through a point cloud encoder that projects geometric features into the LLM’s embedding space, enabling both point cloud captioning and text-conditioned 3D point cloud generation in a single model.

References:

  • Paper: “GPT4Point: A Unified Framework for Point-Language Understanding and Generation (Various, 2024)”

Example

// Create a GPT4Point model for unified point cloud-language understanding
// and 3D point cloud generation from text descriptions
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.TwoDimensional,
    taskType: NeuralNetworkTaskType.ImageClassification,
    inputHeight: 224, inputWidth: 224, inputDepth: 3, outputSize: 512);
            
// ONNX inference mode with pre-trained model
var model = new GPT4Point<double>(architecture, "gpt4point.onnx");
            
// Training mode with native layers
var trainModel = new GPT4Point<double>(architecture, new GPT4PointOptions());

Constructors

Constructor Summary
GPT4Point(NeuralNetworkArchitecture<T> architecture, GPT4PointOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
GPT4Point(NeuralNetworkArchitecture<T> architecture, string modelPath, GPT4PointOptions options)

Properties

Property Summary
int DecoderEmbeddingDim { get; }
int EmbeddingDimension { get; }
int MaxGenerationLength { get; }
int MaxPoints { get; }
int PointChannels { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> EncodeImage(Tensor<T> image)
Tensor<T> GenerateFrom3D(Tensor<T> pointCloud, string prompt) Processes 3D point cloud using GPT4Point’s Point-QFormer alignment approach.
Tensor<T> GenerateFromImage(Tensor<T> image, string prompt) Generates from 2D image using GPT4Point’s Point-QFormer alignment.