README.md

metadata

title: Cairn — Leave, Come Back, Same World
emoji: 🗿
colorFrom: blue
colorTo: gray
sdk: static
pinned: false
license: mit
tags:
  - world-models
  - video-generation
  - long-term-consistency
  - object-permanence
short_description: Leave, come back, same world — and edit it

Cairn — leave, come back, same world

Video world models forget. Turn the camera away from a chair for a few seconds and turn back, and it is a different chair, somewhere else, or gone. The usual fixes give the model more implicit memory — a longer context window, or a compressed latent. Both decay with how long you looked away.

Cairn takes the world out of the weights. Objects live in an explicit external ledger — persistent id, pose, appearance, provenance — written by perception on the model's own generated frames, and read back to coerce generation when the camera returns. A table lookup costs the same whether you looked away for 4 frames or 400.

Over 5 seeds, at 128 frames away, every baseline returns a broken world 0% of the time and Cairn 100%, with a flat 3.6 cm position error.

What this Space does

Three things you can check yourself:

1. Leave & return — drive the camera away from an object for t frames and come back. Cairn OFF and Cairn ON generate from the same scene, the same trajectory and the same seed, so the only difference is the memory.
2. Edit the world — issue move / remove / recolour on an object while it is off screen, turn back, and see whether the instruction stuck. The baselines cannot express the command at all — there is no row to write to.
3. The ledger — the actual table, its transaction log, and a rewind to any earlier frame.

How it runs

Entirely in your browser, via Pyodide — no server, no GPU, no account, nothing uploaded. The first load fetches Python, numpy, scipy and the cairn wheel (~40 MB, cached afterwards); each run then takes a few seconds.

Pyodide has no ffmpeg, so instead of video you get a filmstrip: the opening shot, the last frame before the camera leaves, a frame from while it is away, and the frames after it comes back. For this claim a strip is arguably the better medium — both halves of the comparison sit in one glance.

Not using Gradio here on purpose: gradio-lite imports gradio before installing the page's requirements, and gradio 5.x currently cannot be resolved against huggingface-hub 1.x inside Pyodide. Driving Pyodide directly fixes the ordering and drops gradio's dependency stack from the download.

What you are watching

A surrogate generator that reproduces how autoregressive video drifts (random walk + prior pull + salience decay) — not a real video backbone. That is a deliberate trade: exact ground truth, and a benchmark that runs on a laptop in three minutes, in exchange for not being LTX-Video. The same cairn library wraps a real diffusers video pipeline in one line:

cairn = CairnPipeline.from_pipeline(pipe)

Everything scientific — the generator, perception, the ledger, the forcing path, the metrics — is the same code the benchmark measures, installed here from the same wheel.

Links

• Code, full benchmark, ablation and honest limitations — https://github.com/NagaYu/cairn
• Dataset — https://huggingface.co/datasets/NagaYu/cairn-departure-return

MIT.