Youngjin Shin

Minkyung Kwon, Jinhyeok Choi, Youngjin Shin, Jaeyeong Kim, JongMin Lee, Seungryong Kim

Jiyoung Kim, Youngjin Shin, Siyoon Jin, Dahyun Chung, Jisu Nam, Tongmin Kim, Jongjae Park, Hyeonwoo Kang, Seungryong Kim

Kyungjune Lee, Youngjin Shin, Jungwoo Huh, Sanghoon Lee

An interactive playground reproducing the technique from Olah et al., “Feature Visualization” (Distill, 2017) — explore channels, neurons, layers, logits, probabilities, and joint optimizations.

An interactive 3D playground for DINOv3-L attention — see how each query, key, and value is projected per layer and head, trace a query to its keys and on to their values, and read the full 261×261 attention map with a per-patch heatmap on the image. A second view projects every layer's MLP features to PCA-RGB, showing how each layer's focus shifts with depth.

A reading of six models — REPA, RAE, VGGT and their variants — as one question: where on the depth axis does each place, and mix, its features? With layer-wise linear probes and real per-layer feature maps, it shows low-level detail, high-level semantics, and geometry actively blending inside the networks — including a denoiser's excursion into semantics that the final head returns, dictated by the decoder that reads the latent next.

A close reading of RAEv2's decoder on DINOv3-L: a frozen encoder throws detail away, yet its decoder reconstructs pixels. The post traces how the decoder mirrors the encoder in frequency and representation, why the latent (K) decides what can be reconstructed, where the reconstruction–generation trade-off comes from, and what it means to build a decoder that mirrors on purpose.