Timeline-VLM

Upload images of visual artifacts and the model estimates when they were created by projecting their visual embeddings onto a learned temporal manifold.

📄 Paper  ·  💻 GitHub

How it works

1. Text prompts ("an artifact from the year XXXX") are encoded with CLIP to build a temporal embedding space
2. Kernel PCA (cosine kernel) reduces these to 3-D, revealing a temporal manifold
3. A Bezier curve is fitted through the manifold as a smooth temporal axis
4. Your image is encoded and projected onto that curve, and its position gives the estimated year