DINO-Foresight
DINO-world
VFMF
FlowWM

World modeling in DINO feature space

Can a feature-space world model represent several futures?

Shared recipe
Frozen visual encoder → future features → perception head

Read → through time; ↓ by representation
Orange: point prediction · Green: stochastic generation
Blue “cites” arrows: verified references, pointing to the cited paper. Other arrows: labeled conceptual/benchmark links.

DEC 2024

JUL 2025

DEC 2025

JUN 2026

Compressed features

An extra bottleneck before forecasting.

Full features

Forecast directly in the encoder’s space.

Multi-layer DINO → PCA

Masked transformer
Deterministic regression

One forecast in PCA space

DINO features → VAE

Last-layer DINOv2 features

Cross-attention predictor
Smooth L1 regression

One feature forecast

Autoregressive flow matching
Generate compressed latents

Sample → VAE decode → features

Full pretrained features

Flow matching
Wide head + timestep scheduling

Multiple feature-space futures

Compress to generate?

VFMF: learn a compact space.

FlowWM: adapt the generator.

What if the same past branches?

A point forecast cannot express all alternatives.

RAE · Oct 2025 💻 📦

Making generation in DINO space practical

Semantic features can support high-quality image generation. Adapt the architecture and training to their high dimensionality:

  • Wide, shallow head: avoid a narrow denoising bottleneck.
  • Dimension-aware noise schedule: adjust the noise–signal balance.
  • Noise-augmented decoder: handle imperfect generated latents.

Strong evidence and a practical recipe—not the first feature-space generator.

Frozen Forecasting

Jul 2025 · Frozen Forecasting

Diffusion in frozen features
→ task-specific readouts

FlowWM benchmark baseline:
reimplemented with DINOv3.

Evaluate distributions of trajectories

Compare forecasting across backbones and tasks—not only a single predicted frame.

Public releases · checked 18 Sep 2026
💻 official code · 📦 trained checkpoints
Only available releases get badges.
Backbone weights and local untrained reproductions do not count.

PCA → learned VAE; regression → flowUse full last-layer featuresOne point → conditional distributionRemove the extra VAE bottleneckWide head + dimension-aware timestep scheduleBenchmarked against · DINOv3 reimplementationTrajectory distributionsCitesCites