TLoRAModel<T>
Models & Types · AiDotNet.Diffusion.StyleTransfer
T-LoRA: timestep-dependent low-rank adaptation for customizing a diffusion model from a SINGLE image without overfitting to it.
Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → TLoRAModel<T>
For Beginners
Show a picture generator one photo of your dog and ask it to learn “your dog”, and it will usually memorize that exact photo — same pose, same background, every time. The memorizing happens mainly in the noisy early stage of generation, so this gives the model fewer adjustable directions during that stage and all of them later, when it is only adding detail.
How It Works
Soboleva, Alanov, Kuznetsov and Sobolev, “T-LoRA: Single Image Diffusion Model Customization Without Overfitting” (arXiv:2507.05964). Fine-tuning “frequently suffers from overfitting when training samples are limited, compromising both generalization capability and output diversity”, and this paper takes the hardest version of that: adapting “using just a single concept image”.
PURPOSE CHANGE. This class previously described “temporal LoRA-based style transfer with video-consistent stylization” — applying an artistic style across video frames without flicker. The real T-LoRA is not that. It is single-image CONCEPT personalization: teaching a model a new object from one photo while keeping its diversity. The two share only the letters. The citation this replaced pointed at arXiv:2405.12345, which is a functional-analysis paper on the learning behaviour of paradise fish.
Two innovations, both implemented in TimestepDependentLora.
Constructors
| Constructor | Summary |
|---|---|
TLoRAModel(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, UNetNoisePredictor<T> predictor, StandardVAE<T> vae, IConditioningModule<T> conditioner, int? seed, int? adapterRank) |
Properties
| Property | Summary |
|---|---|
IConditioningModule<T> Conditioner { get; } |
|
INoisePredictor<T> NoisePredictor { get; } |
|
IReadOnlyList<TLoRAAttentionAdapter<T>> Adapters { get; } |
Gets the injected adapters, one per wrapped attention block. |
IVAEModel<T> VAE { get; } |
|
int LatentChannels { get; } |
Methods
| Method | Summary |
|---|---|
ModelMetadata<T> GetModelMetadata() |
|
Tensor<T> PredictNoise(Tensor<T> noisySample, int timestep) |