NARC method slashes diffusion inversion storage by 400x with +3.24 dB PSNR gain
New training-free technique stores just one int8 latent anchor per image, cutting costs dramatically...
A new paper from researchers at an undisclosed institution (authors Yongseong Park, Joeun Kim, HoEun Kim, Young-Sik Kim) tackles the quality-cost trade-off in real-image diffusion inversion. They identify two mechanisms: compression asymmetry and trajectory binding. Compression asymmetry means that storing full-dimensional int8 noise anchors preserves reconstruction quality, whereas low-dimensional subspace summaries often fail even at comparable payloads. Trajectory binding shows that both the matched forward noise anchor and a trained score network are necessary—inversion can't rely solely on algebraic identities.
Building on these insights, the team introduces Noise-Anchored Reverse Correction (NARC), a training-free inversion primitive. NARC stores a single int8 latent anchor and reuses it with a fixed, noise-level-dependent anchor-weight schedule: strong anchoring when reverse trajectory is noise-dominated, then relaxed anchoring as image details emerge. On PIE-Bench++ with Stable Diffusion 1.5, NARC outperforms five non-exact baselines, improving PSNR by +3.24 dB over PnP DirectInv while slashing inversion storage by roughly 400x. The compression asymmetry and anchor specificity also transfer to SDXL at 1024×1024 resolution.
- NARC requires only one int8 latent anchor per image, achieving 400x storage reduction over PnP DirectInv.
- Achieves +3.24 dB PSNR improvement on PIE-Bench++ with Stable Diffusion 1.5, beating five modern baselines.
- Reveals that full-dimensional int8 anchors preserve reconstruction, while subspace summaries collapse—counterintuitive ordering persists across methods.
Why It Matters
Enables high-quality image editing on consumer GPUs with minimal storage, democratizing advanced diffusion inversion.