Why U-Net skip connections matter for denoising

WHAT THIS TESTS: Whether you understand the encoder-decoder trade-off and why lateral skip connections are essential when the output must be pixel-aligned with the input.

A GOOD ANSWER COVERS: A U-Net's encoder progressively downsamples, growing the receptive field and capturing semantic, low-frequency context, but pooling and striding throw away precise spatial location. The decoder upsamples back to full resolution, yet from the bottleneck alone it cannot recover the fine detail that was discarded. Skip connections solve this by concatenating each encoder feature map with the decoder feature map at the same resolution, handing high-frequency spatial information directly across. In the reverse diffusion process the network must output a noise estimate that is pixel-aligned with the noisy input, so preserving exact edges, textures, and object boundaries is critical; skips let the model remove noise without smearing structure. They also create short gradient paths from output to early layers, which stabilizes and speeds training of a deep network.

COMMON WRONG ANSWERS: Confusing the long encoder-to-decoder skips with short residual connections inside a single block; saying skips reintroduce noise rather than spatial detail; claiming the bottleneck alone is sufficient; asserting skips are only for gradient flow and ignoring the spatial-detail role; thinking they merely add parameters.

LIKELY FOLLOW-UPS: Concatenation versus addition for the skip merge, and why concatenation is common here. How does timestep conditioning enter alongside the skips? What happens to image quality if you ablate the skip connections? How do attention blocks coexist with skips at lower resolutions?

ONE CONCRETE EXAMPLE: Denoise a latent where a face has sharp eye and hair boundaries. The encoder's deepest features know it is a face but have lost where each strand of hair sits. The skip from the highest-resolution encoder block carries those crisp boundaries to the decoder, so the predicted noise aligns with the true high-frequency content and the recovered face keeps sharp edges instead of becoming a blurry oval.