PS: 🌀 We recently released spatial-reasoners, a general toolkit to apply SRMs to a wide range of different domains: spatialreasoners.github.io

We find that model hallucination can be drastically reduced by choosing the right configuration, allowing to significantly increase performance in complex reasoning tasks like solving visual Sudoku.

Our Spatial Reasoning Models allow to explore the space between parallel and autoregressive diffusion models with different methods for choosing generation order.

Project Website: geometric-rl.mpi-inf.mpg.de/srm/

Can diffusion models solve visual Sudoku?

If you are at #ICML2025, come to our poster in the Wednesday morning poster session (Poster Session 3 East, Poster 3412) and find out!

@chriswewer.bsky.social

MET3R quantitatively measures 3D consistency between two images via DUSt3R reconstruction and feature comparison. It does not require camera poses.

Code is available for plug-and-play use. We also provide an open source multi-view latent diffusion model for further research!

Project page: geometric-rl.mpi-inf.mpg.de/met3r/

At #CVPR2025 and working on consistency in video and multi-view generative models?

Come and visit our poster on Friday afternoon, where I present 𝗠𝗘𝘁𝟯𝗥: 𝗠𝗲𝗮𝘀𝘂𝗿𝗶𝗻𝗴 𝗠𝘂𝗹𝘁𝗶-𝗩𝗶𝗲𝘄 𝗖𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝗰𝘆 𝗶𝗻 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗲𝗱 𝗜𝗺𝗮𝗴𝗲𝘀

@mohammadasim98.bsky.social @wimmerthomas.bsky.social @mpi-inf.mpg.de @cvml.mpi-inf.mpg.de

We also show that good orders can be predicted by uncertainty, which is crucial for the Sudoku task to be solved well.

Spatial Reasoning Models (SRMs) are a framework to propagate belief over a set of continuous variables (e.g. image patches) with generative denoising models.

It allows to explore the amount of (soft) sequentialization and the order of generation, both having significant impact on reasoning quality.

Can image generators solve visual Sudoku?

Naively, no, with sequentialization and the correct order, they can!

Check out @chriswewer.bsky.social's and Bart's SRM's for details.

Project: geometric-rl.mpi-inf.mpg.de/srm/
Paper: arxiv.org/abs/2502.21075
Code: github.com/Chrixtar/SRM

MET3R measures 3D consistency between two images without camera poses via DUSt3R reconstruction and feature comparison.

Code is available for plug-and-play use. We also provide an open source multi-view latent diffusion model for further research!

Project page: geometric-rl.mpi-inf.mpg.de/met3r/

Hello bluesky-world :)

Introducing 𝗠𝗘𝘁𝟯𝗥: 𝗠𝗲𝗮𝘀𝘂𝗿𝗶𝗻𝗴 𝗠𝘂𝗹𝘁𝗶-𝗩𝗶𝗲𝘄 𝗖𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝗰𝘆 𝗶𝗻 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗲𝗱 𝗜𝗺𝗮𝗴𝗲𝘀.

Lacking 3D consistency in generated images is a limitation of many current multi-view/video/world generative models. To quantitatively measure these inconsistencies, check out Mohammad Asims new work!