LuminaT2XModel<T>

Models & Types · AiDotNet.Diffusion.Video

Lumina-T2X model — transformer-based text-to-any generation (image, video, 3D, audio).

Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → LuminaT2XModel<T>

For Beginners

Lumina-T2X is a unified model for multiple generation tasks:

Key characteristics:

  • Single architecture generates images, videos, 3D, and audio
  • Flag-DiT: improved DiT with flow matching
  • Resolution-aware encoding: handles any aspect ratio
  • Gemma text encoder for multilingual prompts
  • Scalable from 0.6B to 7B parameters

Use Lumina-T2X when you need:

  • Multi-modal generation from text
  • Flexible resolution/aspect ratio support
  • Research into unified generation

How It Works

Lumina-T2X is a unified transformer framework for generating multiple modalities from text: images, videos, 3D objects, and audio. It uses a Flag-DiT backbone with resolution-aware positional encoding.

Technical specifications:

  • Architecture: Flag-DiT with flow matching
  • Text encoder: Gemma-2B (2048-dim)
  • Resolution: up to 2048x2048
  • Latent channels: 4, 8x downsampling
  • Flow matching with velocity prediction

Reference: Gao et al., “Lumina-T2X: Transforming Text into Any Modality”, 2024

Example

var options = new DiffusionModelOptions<float> { LatentChannels = 4, DefaultInferenceSteps = 30 };
            
var trainX = Tensor<float>.CreateRandom(2, 4, 32, 32);
var trainY = Tensor<float>.CreateRandom(2, 4, 32, 32);
            
var result = new AiModelBuilder<float, Tensor<float>, Tensor<float>>()
    .ConfigureModel(new LuminaT2XModel<float>(options: options))
    .Build(trainX, trainY);
            
var noise = Tensor<float>.CreateRandom(new[] { 1, 4, 128, 128 });
var generated = result.Predict(noise);

Constructors

Constructor Summary
LuminaT2XModel(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, DiTNoisePredictor<T> dit, StandardVAE<T> vae, IConditioningModule<T> conditioner, int? seed)

Properties

Property Summary
IConditioningModule<T> Conditioner { get; }
INoisePredictor<T> NoisePredictor { get; }
IVAEModel<T> VAE { get; }
int LatentChannels { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> GenerateFromText(string prompt, string negativePrompt, int width, int height, int numInferenceSteps, double? guidanceScale, int? seed)

Fields

Field Summary
int DefaultHeight
int DefaultWidth