Puffin-World

Supplementary Results

More qualitative results across camera-to-world understanding, camera-controllable generation, and native 3D world modeling.

Back to Puffin-World

Camera-to-World Understanding

Perspective fields across diverse datasets

Puffin-World predicts roll, pitch, and vertical field of view from each image and converts them into latitude maps and gravity fields. The results remain stable across indoor, outdoor, synthetic, and real-world scenes.

Perspective field visualizations across multiple public datasets — Perspective field visualizations on different datasets, with camera parameters predicted by Puffin-World.

Camera-Controllable Generation

Free-viewpoint simulation across scenes

Explicit camera conditions produce visible yet physically plausible changes in viewpoint, horizon, and spatial composition while preserving realistic appearance and prompt consistency.

Camera-controllable text-to-image generation across indoor and outdoor scenes — Camera-controllable text-to-image generation for free-viewpoint spatial simulation.

Rotation Control

Explicit roll, pitch, and yaw trajectories

Puffin-World follows large rotational controls along both single-pass and recursive generation paths while retaining the scene content and a coherent gravity-aligned physical frame.

3D World Generation

Long trajectories and compound camera actions

Additional examples cover long-horizon motion, large viewpoint changes, and combined camera actions. Generated views preserve semantic content and coherent spatial structure across extended trajectories.

Additional Puffin-World generation results under long and compound camera trajectories — Additional qualitative results of 3D world generation under diverse trajectories and compound camera actions.

Physics Propagation

Ground long trajectories in a consistent physical frame

Physics propagation transfers the absolute physical state from the reference view across relative camera motion, reducing horizon drift and improving both visual fidelity and camera alignment.

Physics Propagation

More consistent long trajectories

Anchoring all views to a shared gravity-aligned frame improves reconstruction quality and camera alignment across challenging motions.

PSNR19.87 → 20.54 SSIM0.62 → 0.67 LPIPS ↓0.23 → 0.21 Roll error ↓1.94° → 1.76° Pitch error ↓5.14° → 4.75°

Qualitative ablation comparing generated camera trajectories with and without physics propagation using up-vector error — Qualitative ablation of physics propagation. Each pair compares generation with and without PP using up-vector error; lower values and greener vectors indicate better gravity alignment.

Scaling with Native 3D World States

Appearance, geometry, physics, and reconstruction

Puffin-World jointly predicts visual appearance, depth-based geometry, and gravity-aware physics. The generated observations can then be consolidated into clear and spatially coherent 3D reconstructions.

Native appearance, geometry, and physics states with 3D reconstruction — More qualitative results of 3D world modeling. Each example shows native appearance, geometry, and physics states together with the reconstructed 3D scene.

Puffin-World

Return to the main project page

Explore the unified framework, interactive 3D viewer, quantitative evaluation, datasets, code, and models.

Back to Puffin-World