Ideogram3Model<T>

Models & Types · AiDotNet.Diffusion.TextToImage

Ideogram 3 model for text-to-image generation with superior text rendering.

Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → Ideogram3Model<T>

For Beginners

Most image generators struggle to write readable text in images. Ideogram 3 excels at this — it can generate signs, logos, posters, and business cards with correctly spelled, properly formatted text.

How Ideogram 3 works:

  1. Text prompt is parsed to identify embedded text requests (e.g., “a sign that says ‘Hello’”)
  2. A text layout predictor determines optimal placement, font size, and orientation
  3. The SiT diffusion backbone generates the image with text-aware conditioning
  4. OCR-in-the-loop training ensures spelling accuracy

Key characteristics:

  • Text layout prediction module for accurate text placement
  • OCR-in-the-loop training for spelling accuracy
  • SiT (Scalable Interpolant Transformer) backbone
  • Handles multi-line text, different fonts, and curved text
  • 16 latent channels

Advantages:

  • Best-in-class text rendering accuracy
  • Correct spelling even for complex words
  • Natural text placement and sizing
  • Handles various text styles (signs, posters, book covers)
  • Good overall image quality beyond text

Limitations:

  • API-only access
  • Text rendering adds inference overhead
  • Less control over exact font/style compared to dedicated design tools

How It Works

Ideogram 3 specializes in generating images with accurate, legible text rendering. It uses a specialized text layout prediction module alongside a SiT diffusion backbone to ensure rendered text is spelled correctly, properly formatted, and placed naturally within the generated scene.

Technical specifications:

  • Architecture: SiT with text layout prediction
  • Text layout module: Predicts bounding boxes, font size, orientation
  • OCR-in-the-loop: Training-time OCR validation for spelling
  • Backbone: SiT, hidden 2048, 24 layers
  • VAE: 16 latent channels
  • Default: 30 steps, guidance scale 7.5
  • Resolution: 1024x1024 default

Reference: Ideogram Inc., “Ideogram 3”, 2024

Example

// Create Ideogram 3
var ideogram = new Ideogram3Model<float>();
            
// Generate an image with text
var image = ideogram.GenerateFromText(
    prompt: "A vintage movie poster that says 'The Last Journey' in art deco style",
    width: 1024,
    height: 1024,
    numInferenceSteps: 30,
    guidanceScale: 7.5,
    seed: 42);

Constructors

Constructor Summary
Ideogram3Model(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, SiTPredictor<T> predictor, StandardVAE<T> vae, IConditioningModule<T> conditioner, int? seed) Initializes a new instance of Ideogram3Model.

Properties

Property Summary
IConditioningModule<T> Conditioner { get; }
INoisePredictor<T> NoisePredictor { get; }
IVAEModel<T> VAE { get; }
bool SupportsTextRendering { get; } Gets whether this model has specialized text rendering capabilities.
bool UsesOCRTraining { get; } Gets whether this model uses OCR-in-the-loop training for spelling accuracy.
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)
Tensor<T> ImageToImage(Tensor<T> inputImage, string prompt, string negativePrompt, double strength, int numInferenceSteps, double? guidanceScale, int? seed)

Fields

Field Summary
int DefaultHeight Default image height for Ideogram 3 (1024x1024).
int DefaultWidth Default image width for Ideogram 3 (1024x1024).