U-Net
U-Net is an encoder–decoder CNN with skip connections from downsampling to upsampling paths, designed so fine spatial detail survives compression.
Definition
Ronneberger et al. built a U-shaped conv net for biomedical segmentation: contract, bottleneck, expand, and concatenate encoder feature maps into the decoder. Those skips preserve edges. The same skeleton is the default denoiser in pixel-space diffusion.
Why it matters
If a task is dense prediction — masks, depth, noise residual — U-Net is the inductive bias you try first. Diffusion image models did not invent it; they borrowed it.
Case
Cell-tracking challenges in 2015 were won by U-Nets trained on few annotated slides. A decade later, ε-prediction U-Nets generate ImageNet-scale pictures by the same skip logic.
How to read it
Check skip alignment (channel count, resolution) and whether the bottleneck is too small for the semantics you need. Broken skips show up as blurry edges, not as a NaN.
Key papers
Landmark references with real publisher or arXiv links — not token-overlap noise.
- U-Net: Convolutional Networks for Biomedical Image SegmentationRonneberger, Fischer, Brox · 2015 · MICCAIThe original U-Net.