timestep embedding

**Timestep embedding** is the **numeric representation of diffusion step index or noise level used to condition denoiser behavior** - it tells the network how much corruption is present so each layer can apply the right denoising operation. **What Is Timestep embedding?** - **Definition**: Encodes time or sigma values into feature vectors, often with sinusoidal functions and MLP projection. - **Injection**: Added into residual blocks so denoising behavior changes across noise levels. - **Continuous Support**: Can represent fractional timesteps for advanced ODE samplers. - **Compatibility**: Works jointly with text conditioning and other control embeddings. **Why Timestep embedding Matters** - **Denoising Accuracy**: Correct time encoding is required for stable predictions across the noise trajectory. - **Sampler Fidelity**: Good timestep conditioning improves behavior under reduced step schedules. - **Transferability**: Consistent embedding design helps checkpoint portability across inference stacks. - **Guidance Stability**: Weak timestep signals can amplify artifacts under strong guidance. - **Optimization**: Embedding architecture choices influence training speed and convergence quality. **How It Is Used in Practice** - **Scaling**: Normalize timestep ranges consistently between training and inference code paths. - **Ablation**: Compare sinusoidal plus MLP against learned embeddings for target domains. - **Validation**: Test sampler families that use nonuniform steps to verify robust interpolation behavior. Timestep embedding is **a required conditioning signal for accurate diffusion denoising** - timestep embedding quality directly affects stability, fidelity, and sampler interoperability.