Slot-RAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders

Name: Slot-RAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders
Uploaded: 2026-10-03T12:24:47.849116+00:00
Duration: 1013 s
Description: SlotRAE introduces a simpler framework to understand and generate real-world scenes by operating directly within vision foundation models' semantic feature space, like DINOv3. It unifies scene decomposition (using object slots) with image generation via a Diffusion Transformer. This approach avoids the bottlenecks of traditional VAEs while achieving state-of-the-art results on COCO.

SlotRAE introduces a simpler framework to understand and generate real-world scenes by operating directly within vision foundation models' semantic feature space, like DINOv3. It unifies scene decomposition (using object slots) with image generation via a Diffusion Transformer. This approach avoids the bottlenecks of traditional VAEs while achieving state-of-the-art results on COCO.

Open the full episode