VILAU<T>

Models & Types · AiDotNet.VisionLanguage.InstructionTuned

VILA-U: unified VLM with autoregressive visual generation and understanding.

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

Implements: IInstructionTunedVLM<T>

For Beginners

VILA-U extends VILA to not just understand images but also generate them. Most vision-language models are one-way — they look at images and produce text. VILA-U works both ways: it can analyze an image and describe it, or take a text description and generate an image. It does this by converting images into discrete tokens (like a visual vocabulary) that the language model can both read and write. This unified approach means a single model handles image understanding, text generation, and image generation without needing separate specialized models for each task. Default values follow the original paper settings.

How It Works

VILA-U (NVIDIA, 2024) unifies visual understanding and visual generation in a single autoregressive model. Unlike most VLMs that can only understand images and generate text, VILA-U can also generate images autoregressively by predicting visual tokens one at a time. It uses a discrete visual tokenizer to convert images into a sequence of tokens that the language model can both consume (for understanding) and produce (for generation), creating a truly unified visual-linguistic model.

References:

  • Paper: “VILA-U: a Unified Foundation Model Integrating Visual Understanding and Generation” (NVIDIA, 2024)

Example

// Create a VILA-U model for unified visual understanding and generation
// using autoregressive visual token prediction
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 VILAU<double>(architecture, "vilau.onnx");
            
// Training mode with native layers
var trainModel = new VILAU<double>(architecture, new VILAUOptions());

Constructors

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

Properties

Property Summary
int DecoderEmbeddingDim { get; }
int EmbeddingDimension { get; }
int MaxGenerationLength { get; }
string LanguageModelName { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Chat(Tensor<T> image, IEnumerable<ValueTuple<string, string>> conversationHistory, string userMessage)
Tensor<T> EncodeImage(Tensor<T> image)
Tensor<T> GenerateFromImage(Tensor<T> image, string prompt) Generates text using VILA-U’s unified understanding and generation architecture.