Diffusion & generation

Timestep embedding

Tells the network how noisy the input currently is, so one network can handle every denoising stage.

First page of Denoising Diffusion Probabilistic ModelsPopularized inJun 2020Denoising Diffusion Probabilistic ModelsHo et al. · arXiv 2006.11239 ↗
Timestep embedding: telling the net how noisyt samples a stack of sinusoids · the pattern becomes a conditioning vectortinjected into every block

The same network weights must denoise both nearly-pure noise and nearly-finished images. The current timestep is therefore encoded (usually sinusoidally, then through a small MLP) and injected into every block, letting the network condition its behavior on how far along the denoising process is.

Adoption over time

Share of new models that have included timestep embeddings over time.

6%202220232024202520266%20222023202420252026

See it in real models

Open any of these on hfviewer to find this block in the interactive architecture graph.

ResembleAI/chatterbox architecture graphResembleAI/chatterboxtext-to-speech · ↓ 2.5M · ♡ 2kOpen in visualizer google/diffusiongemma-26B-A4B-it architecture graphgoogle/diffusiongemma-26B-A4B-itimage-text-to-text · ↓ 2.0M · ♡ 1kOpen in visualizer nvidia/diffusiongemma-26B-A4B-it-NVFP4 architecture graphnvidia/diffusiongemma-26B-A4B-it-NVFP4text-generation · ↓ 1.7M · ♡ 115Open in visualizer Tongyi-MAI/Z-Image-Turbo architecture graphTongyi-MAI/Z-Image-Turbotext-to-image · ↓ 1.1M · ♡ 5kOpen in visualizer openbmb/VoxCPM2 architecture graphopenbmb/VoxCPM2text-to-speech · ↓ 900k · ♡ 2kOpen in visualizer microsoft/VibeVoice-ASR architecture graphmicrosoft/VibeVoice-ASRautomatic-speech-recognition · ↓ 695k · ♡ 1kOpen in visualizer microsoft/VibeVoice-Realtime-0.5B architecture graphmicrosoft/VibeVoice-Realtime-0.5Btext-to-speech · ↓ 657k · ♡ 1kOpen in visualizer krea/Krea-2-Raw architecture graphkrea/Krea-2-Rawtext-to-image · ↓ 133k · ♡ 433Open in visualizer ByteDance/SDXL-Lightning architecture graphByteDance/SDXL-Lightningtext-to-image · ↓ 104k · ♡ 2kOpen in visualizer xinsir/controlnet-union-sdxl-1.0 architecture graphxinsir/controlnet-union-sdxl-1.0text-to-image · ↓ 100k · ♡ 2kOpen in visualizer microsoft/VibeVoice-1.5B architecture graphmicrosoft/VibeVoice-1.5Btext-to-speech · ↓ 75k · ♡ 2kOpen in visualizer Qwen/Qwen-Image-2512 architecture graphQwen/Qwen-Image-2512text-to-image · ↓ 73k · ♡ 921Open in visualizer

Browse all 103 models with this in the catalog →

Related concepts

VAE (latent autoencoder) DiT block (diffusion transformer) Denoising loop / scheduler

hfviewer renders the full architecture of 3,100+ Hugging Face models as interactive graphs. Hover any block to see what it does, with this model’s real numbers.

Browse all model graphs →