REIMPLEMENTATION COURSE · CVPR 2025
Understand CUT3R by building the smallest version that can fail honestly.
The goal is not to copy 793M parameters. The goal is to make the recurrent world-state idea feel inevitable: observe → update memory → read out geometry → test what memory actually contains.
THE LOOP
RGB frame
Iₜ
→Iₜ
image tokens
Fₜ
Fₜ
↘ state read / update ↙
persistent state
Sₜ₋₁ → Sₜ
→Sₜ₋₁ → Sₜ
pointmap + pose
Xₜ, Pₜ
Xₜ, Pₜ
Sₜ, Rₜ = fθ(Sₜ₋₁, Fₜ)
01Implement the invariant first
A fixed-size state must carry useful information across arbitrarily many observations.
02Make geometry explicit
Coordinate frames, pointmaps and poses are easier to debug than a mysterious “3D loss”.
03Ablate memory aggressively
If resetting or shuffling state does not hurt, you did not build a recurrent world model.
YOUR RESEARCH LOG
One running notebook
Write hypotheses, broken assumptions and measurements here. It is stored locally in this browser.