Skip to content

Dataset · KangLiao

Multimodal dataset with camera parameter binding

KangLiao/Puffin-4M

A large-scale image-text dataset with camera geometry for spatial AI.

What's inside

Contains images and texts combined with camera parameters to train spatial understanding and generation. Data volume ranges from 1 to 10 billion records (4M indicated in the name). Created by the Thinking with Camera project authors specifically to address data scarcity at the intersection of vision, language, and camera geometry.

What you can do with it

  • Train multimodal models to understand spatial geometry and camera angles
  • Fine-tune image and 3D scene generators with precise camera position control
  • Benchmark spatial intelligence in VLM models
Tasks
Text → image, Image → text, Image → 3d, Image → image
Size
1B–10B rows
Downloads this month
15.6K
Likes
35

How to load it

from datasets import load_dataset

ds = load_dataset("KangLiao/Puffin-4M")