NTU Researchers Release Puffin-World to Put Camera Geometry Inside a World Model

Nanyang Technological University researchers have released Puffin-World, a world model built to generate views from specified camera movements while retaining explicit representations of gravity, camera orientation and depth. The release gives other labs the model, code and data to test that premise; for now, its strongest reported results come from the team’s own evaluation materials rather than independent replication.

A scene model, not just a sequence of frames

Puffin-World jointly predicts three kinds of state: physics, geometry and appearance. Those states include a gravity field and latitude map for uprightness and viewing-ray elevation, depth for geometry, and RGB images or sequences for appearance. The researchers’ aim is to keep the camera’s relationship to a scene available to the model rather than asking pixels alone to preserve spatial structure.

Its Omni-Camera conditioning assigns each pixel nine camera channels. Three describe an absolute field with an up vector and latitude angle; six describe relative ray origin and direction. The combination is intended to retain a global gravity reference as the viewpoint changes continuously.

Starting from an image or text prompt, the model can generate views along paths involving translation, roll, pitch, yaw and compound movement. It can also estimate camera parameters from an uncalibrated image, produce RGB and depth views together, and consolidate those outputs into a 3D point cloud.

15 millionVision-language-camera triplets

Puffin-Cam-15M was rendered from roughly 900,000 panoramas.

1 millionCamera trajectories

Puffin-Traj-1M covers camera movement patterns including vertical motion, rotations, 360-degree exploration and compound movements.

44.5 millionAnnotated public-dataset images

The team used Puffin-World to add roll, pitch and vertical field-of-view information across 28 public datasets.

The data is designed around viewpoint variation

The Puffin-16M release combines its panorama-derived examples with trajectory data meant to expose the model to controlled changes in where a camera is and how it is pointed. The researchers also say the 44.5 million added camera annotations can help study viewpoint bias and build training sets with more varied camera configurations.

What the public package includes

• Model checkpoints for running and inspecting the released systems.
• A public code repository and the Puffin-16M training dataset.
• Supplementary evaluation materials accompanying the release.

A testable claim, not a settled result

The model card reports the best median errors, including ties, in all 12 cited camera-understanding comparisons across Stanford2D3D, MegaDepth, TartanAir and LaMAR. It also reports first-place PSNR and LPIPS image-quality results on RealEstate10K. Those are useful initial signals, but they are author-reported figures and have not been independently replicated.

Puffin-World is an academic research distribution, not a commercial service. The material question now is whether outside researchers can reproduce the reported gains and whether the model’s geometry remains stable beyond the curated demonstrations used to present it.