Thinking with Camera: A Unified Multimodal Model for Camera-Centric Understanding and Generation

TL;DR

Puffin presents a unified camera-centric multimodal framework that treats camera geometry as a first-class modality, enabling joint camera understanding and camera-controllable generation across arbitrary viewpoints. It combines a geometry-aligned vision encoder, a large language model, and a diffusion generator, all connected through a dedicated thinking with camera paradigm and a pixel-wise camera map latent to capture fine-grained spatial cues. The authors construct Puffin-4M, a four-million triplet dataset of vision-language-camera data, and demonstrate state-of-the-art or competitive performance on camera understanding and cross-view generation, with instruction-tuned variants supporting spatial imagination, world exploration, and photographic guidance. This work advances spatial intelligence by aligning geometric reasoning with linguistic and perceptual priors, enabling practical tasks such as 3D object insertion and cross-view content creation, while providing open benchmarks and code for future research.

Abstract

Camera-centric understanding and generation are two cornerstones of spatial intelligence, yet they are typically studied in isolation. We present Puffin, a unified camera-centric multimodal model that extends spatial awareness along the camera dimension. Puffin integrates language regression and diffusion-based generation to interpret and create scenes from arbitrary viewpoints. To bridge the modality gap between cameras and vision-language, we introduce a novel paradigm that treats camera as language, enabling thinking with camera. This guides the model to align spatially grounded visual cues with photographic terminology while reasoning across geometric context. Puffin is trained on Puffin-4M, a large-scale dataset of 4 million vision-language-camera triplets. We incorporate both global camera parameters and pixel-wise camera maps, yielding flexible and reliable spatial generation. Experiments demonstrate Puffin superior performance over specialized models for camera-centric generation and understanding. With instruction tuning, Puffin generalizes to diverse cross-view tasks such as spatial imagination, world exploration, and photography guidance. We will release the code, models, dataset pipeline, and benchmark to advance multimodal spatial intelligence research.

Figures (19)

• Figure 1: Illustration of the versatile capabilities of our Puffin model. It unifies camera-centric generation (a) and understanding (b), supports the thinking mode (c), and enables diverse applications (d).
• Figure 2: Overview of the proposed Puffin. It jointly learns the camera-centric understanding and generation tasks in a unified multimodal framework. The elements bounded with dotted boundaries represent the cross-view understanding and generation during instruction tuning, such as spatial imagination and world exploration.
• Figure 3: Methods for learning camera geometry. (Left) Previous classical and learning-based methods focused on extracting or learning representations such as geometric structures or semantic features (with confidence). (Right) We introduce the notion of thinking with camera through LMMs. It first decouples the camera parameters across geometric context, establishing connections between spatially grounded visual cues (highlighted in the masked regions) and professional photographic terms. The camera parameters are then predicted within the <answer></answer> tag through this spatial reasoning process <think></think>.
• Figure 4: Overview of the proposed Puffin-4M dataset. It consists of 4 million vision-language-camera triplets under various scenarios and camera configurations. We mark the sample images with different colors, each denoting a different variant of the camera configurations.
• Figure 5: Pipeline of the dataset construction. P2T denotes the mapping from the numerical camera parameters to the professional photographic terms. For clarity, we omit the orientations "clockwise" and "counterclockwise" of the Dutch angle in photographic terms.