Probabilistic Prediction
A single graphical model expresses conditional distributions between any subset of tokens — an LLM-like prompting interface for video.
psi
Probabilistic Structure Integration
What is PSI?
A richly promptable world model that lets us discover the underlying structure of the world.
PSI is a system organized around a three-step iterative cycle. At each turn of the cycle it becomes a sharper predictor and a richer interface onto the world it models.
Structure Extraction
Zero-shot causal inference pulls out low-dimensional properties of the world: depth, flow, segments, control signals.
Integration
Extracted structures are re-encoded as new token types and folded back into training, so each cycle compounds.
Inference engine
Prompt it like an equation.
Drag tokens into the inputs and output slots. PSI treats every modality — RGB frames, depth, dense features, control pokes — as a token, so any subset can be a query and any other subset the answer.
... -> ...
Quick Start
Run PSI in a few lines.
Run any causal inference with a unified interface.
quickstart.py
from PIL import Image
from transformers import AutoModel
predictor = AutoModel.from_pretrained(
"StanfordNeuroAILab/psi0_5",
trust_remote_code=True,
device="cuda",
)
rgb0 = Image.open("frame0.png")
rgb1 = predictor.generate("rgb0->rgb1", rgb0=rgb0)
rgb1.save("frame1.png")