NTU's AnyBokeh edits bokeh from any focus and aperture on a single photo, accepted at NeurIPS
ccloy · x · 2026-10-30
Researchers at NTU's S-Lab (Xinyu Hou et al., advised by Chen Change Loy) present AnyBokeh, a physics-guided any-to-any bokeh editing framework accepted at NeurIPS 2026.
Key points:
- Works on photos captured under arbitrary focus and aperture — no all-in-focus input or reconstruction pipeline needed; existing blur is read as an "optical fingerprint" instead of discarded
- Estimates source blur state via a signed circle-of-confusion map and disparity map, then transfers the source's optical characteristics to any target focus/aperture setting
- A generative editor conditioned on source and target CoC maps performs spatially adaptive deblurring, preservation, and defocus rendering
- Trained with physics supervision on a new high-fidelity synthetic dataset with accurate depth, focus distance, and full EXIF metadata
Paper, code, and models are already available; dataset coming soon.
More from Multimodal
- Midjourney sref code 1461958802 goes viral for eerie green-code retro sci-fi aesthetic — egeberkina · 2026-10-30
- Prompt share: turn any character into a Pixar-style 3D render — azed_ai · 2026-10-30
- CapCut GPT in ChatGPT turns a prompt into a finished short video with script and voiceover in 5 minutes — thisguyknowsai · 2026-10-30
- Peking University's NeuOctree predicts ray-surface intersections for direct ray-traced rendering — ssh4net · 2026-10-30
- Two-person AI fight choreography on an RTX 3060: MiniMax H3 with first-frame, 18 storyboard frames and 3 LoRAs — JinnTsui · 2026-10-30
- MiniMax H3 video-to-video on 8GB VRAM: 832x640 renders in ~19 minutes on RTX 4070 — big-boss_97 · 2026-10-30