📢📢📢 Introducing 𝐒𝐨𝐥𝐯𝟑𝐑 Can geometric foundation models move beyond fixed-set prediction and keep solving like classical SfM? We turn a looped Transformer into a reusable 𝐥𝐞𝐚𝐫𝐧𝐞𝐝 𝐠𝐞𝐨𝐦𝐞𝐭𝐫𝐢𝐜 𝐬𝐨𝐥𝐯𝐞𝐫 🌐 Project: 🧵👇

One solver, every setting. Whether images arrive all at once (offline), frame by frame (streaming), or in arbitrary order, the same Solv3R handles them all with same logic—without task-specific training.

How can one model handle all these settings? We turn a looped Transformer into a multi-view solver. Solv3R applies the same learned geometric update at every call, independent of the iteration count, allowing it to continue solving from any state.

What was solved. Stays solved. A solver shouldn't lose a good reconstruction just because it keeps running. Solv3R learns to preserve accurate states, remaining stable even beyond its training horizon.

New observations also correct the old ones. When new views arrive, Solv3R continues from its current state, jointly refining old and new views. New observations can improve the reconstruction already built.

One frame out at every call (pipelining) Solv3R can pipeline frames at different refinement stages, rather than re-solving the entire window for each new frame. With 16 calls, this gives 16× solver throughput at the same latency.

One cache for every call. Solved views serve as reusable geometric context. Unlike models that need a separate KV cache for each layer or refinement step, Solv3R reuses the same cache across all solver calls.

And finally the solv3r arXiv link. Was supposed to be yesterday at 5pm, paper disappeared from the list without warning, then reappeared without warning... 👀👀

the stopping rule fires when the state stops moving. but your failure case also stops moving, at the wrong fixed point. is there anything in the dynamics that separates converged from stuck, or is that still open?

Awesome, I was looking for someone for doing this. Any chance to have code anytime soon (potential baseline for a potential CVPR submission)?

Let me check if I can give you private early access. Is it fine if it's not super-polished? For a baseline I am sure a coding agent can find its way around.

Yeah, we don't need polished.

Wow!

When code+model? Wow waiting

About a week, depending how much time Jeong Gi will sleep 😅

this looks crazy, looking into this, hope its metric