Puffin-World cuts gravity error to 1.32° with 3D world states

Puffin-World, a new unified multimodal model from Nanyang Technological University's S-Lab and collaborators including the University of Michigan and Beijing Jiaotong University, reports a gravity-orientation error of 1.32° on its own camera-controllable generation benchmark, against 28.31° for FLUX.2-dev and 4.92° for the team's earlier Puffin.

The paper posted to Hugging Face's papers hub frames the design around what it calls 'native 3D world states': the framework jointly models 'physics (gravity field and latitude), geometry (depth), and appearance (image),' tied together by a 9-channel Omni-Camera representation that fuses an absolute camera field with a relative ray field.

The authors write that Puffin-World 'integrates physical understanding, spatial simulation, and 3D world generation and reconstruction without relying on external offline modules.' A separate physics-propagation step carries the reference-frame gravity direction across future views, which the paper credits with 'more stable horizons, better scene uprightness, improved perceptual quality.'

Numbers beyond the gravity headline are more modest. On RealEstate10K, PSNR comes in at 17.22 and LPIPS at 0.318. The Puffin-Cam-Bench used for the 1.32° figure is the team's own construction, and the abstract does not put FLUX.2-dev or the older Puffin on the RealEstate10K table for a matched comparison. It lands amid a run of camera- and world-model work we've tracked, including World Labs' Atlas omni world model.

The training corpus is the other unlock. The team assembled Puffin-16M: 15 million vision-language-camera triplets plus 1 million trajectories, drawn from a pool of 44.5 million images across 28 public datasets that were re-annotated with absolute camera parameters. Code, models, and datasets are released. It arrives the same week Nvidia signed a definitive $12.9B deal to acquire Hugging Face, which now hosts a growing share of the open multimodal frontier.