RAESR Does Super-Resolution in Frozen DINOv3 Latent Space at 37ms per 512px Image
RTP-LLM · hf · 2026-10-15
Researchers argue reconstruction-oriented latent/pixel spaces are the wrong substrate for real-world super-resolution, and propose using the fused latent space of 23 frozen DINOv3-L layers instead, where degraded embeddings sit near the natural-image manifold and semantics guide faithful restoration. A 415M decoder trained with reconstruction and adversarial objectives (RAESR) achieves the best fidelity–perception trade-off among SOTA adversarial and diffusion restorers on RealSR, DRealSR, LSDIR and DIV2K-Val, at 37ms per 512×512 image on one H20 GPU; swapping in a VAE latent loses on every metric.
More from Multimodal
- Claymation-style AI video of a carnival ring toss charms Reddit — Ok-Salamander-3477 · 2026-10-15
- AI portrait thread shows photoreal red-haired fashion shots rivaling studio photography — LudovicCreator · 2026-10-15
- AI video fumbles a handshake: hands still break in generated clips — misovalko · 2026-10-15
- BigCell Generates Coherent Gigapixel Whole-Slide Pathology Images, Boosting Few-Shot Classification 18% — StonyBrook-CVLab · 2026-10-15
- Austin AI Film Festival screens 700+ submissions across 11 categories this week — lmoroney · 2026-10-15
- Fish Audio's Drama-3 voice-cloning Space trends on Hugging Face — fishaudio · 2026-10-15