Weekly Research Digest — 2026-W17

This week clusters around two threads: generalist vision via image generation (Vision Banana) and 3D/4D reconstruction stacks closing the loop with diffusion priors (AnyRecon, Vista4D, JOSH).

Threads

• Generative perception: image generators as a unified interface for segmentation and metric depth.
• 3D Gaussian Splatting reaching consumer capture pipelines.
• 4D reconstruction conditioned on geometry-aware video diffusion.

Linked papers: arxiv.