RAEv2

Summary

RAEv2 is a representation-autoencoder recipe that aggregates multiple pretrained vision-encoder layers, combines RAE with REPA, and reuses the REPA head for single-forward-pass internal guidance.

Role In The Wiki

RAEv2 is the current source for treating layer aggregation as a practical latent-interface knob for unified understanding/generation and for action-conditioned navigation video rollouts.

Official Artifacts

• Project page: https://raev2.github.io/
• Code: https://github.com/nanovisionx/RAEv2
• Hugging Face collection: https://huggingface.co/collections/nyu-visionx/raev2
• Official weights: https://huggingface.co/nyu-visionx/RAEv2-models
• Official data: https://huggingface.co/datasets/nanovisionx/RAEv2-data

Evidence

• Improved Baselines with Representation Autoencoders

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in raev2-2026 rather than duplicating verdict rows here.

At the entity level, RAEv2 should stay as the object card for the method, code, models, and dataset artifacts. The source page carries the evidence ledger, the X discussion caveats, and the slot-level verdicts.

Related Pages

• Intermediate-Layer Representations
• Vision Foundation Models
• Latent Tokenization
• Latent-Space Predictive Learning
• Unified Multimodal Models
• World Models
• Robotics Time-Series Modeling
• MoDA