Super-Resolution in The Right Latent Space: A Frozen Vision-Foundation Substrate
RAESR restores images in a frozen DINOv3-L latent space, leading reported super-resolution quality and speed.
The paper argues that reconstruction-oriented latent spaces and pixel space are poor substrates for real-world super-resolution because degraded images sit far from the natural-image manifold. RAESR instead uses 23 fused layers of a frozen DINOv3-L and trains a 415M decoder with reconstruction and adversarial objectives. It reports the best fidelity-perception trade-off among adversarial and diffusion restorers on RealSR, DRealSR, LSDIR, and DIV2K-Val, at 37 ms per 512 by 512 image on one H20 GPU. Replacing the substrate with a VAE latent under the same recipe loses on every metric.
- Uses 23 fused layers of frozen DINOv3-L as the restoration substrate
- A 415M decoder maps degraded embeddings back in one pass
- Best fidelity-perception trade-off on RealSR, DRealSR, LSDIR, and DIV2K-Val
- Runs in 37 ms per 512 by 512 image on one H20 GPU
- An identical recipe on a VAE latent loses on every metric
From the original article
1 min read at huggingface.co
In an image latent space, the embeddings of high-resolution, natural, and sharp images form a manifold. Degradation of high-resolution images pushes their embeddings off this manifold.
By Wanzhou Lei, Cuifeng Shen, Yanjin He · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.