Parametric Shadow Control for Portrait Generation in Text-to-Image Diffusion Models
ICCV 2025
Put a shadow knob on a text-to-image model — direction, softness and shape, with no retraining. Move your cursor across to walk the light around her face.
Haoming Cai
/haʊˈmɪŋ tsaɪ/ · 蔡昊明
Generative Media
Computational Imaging
Fourth-year CS PhD student at the University of Maryland, advised by Christopher Metzler. I work on generative media — image and video models — teaching them to respect light, geometry, and the camera.
Previously at X-Pixel (2020–2022) with Chao Dong and Jinjin Gu; BEng from The Chinese University of Hong Kong, Shenzhen. Always happy to talk research or early-career questions.
Selected Work
Both figures are interactive — move your cursor across them.
Large-Scale Light Field Synthesis from Videos Enables Geometrically Consistent Bokeh Editing
ECCV 2026
Render geometrically consistent bokeh on scenes that defeat other methods — here, a toy behind a transparent lid. Move your cursor across to pull focus through it.
News
-
Our work on vision–language robustness under adverse imaging was accepted to CoLM 2026. Congratulations to Tianfu and Mingyang.
-
Large-Scale Light Field Synthesis from Videos was accepted to ECCV 2026.
-
Back at the Adobe NextCam team for a second summer internship.
-
Parametric Shadow Control was accepted to ICCV 2025.
-
Named an Outstanding Reviewer for CVPR 2025.
-
Joined the Adobe NextCam team (led by Marc Levoy) as a summer intern, mentored by Shumian Xin and Zhoutong Zhang.
-
Two papers accepted to CVPR 2025. Congratulations to Tianfu, Mingyang, and Jingxi.
All Publications Google Scholar →
* denotes equal contribution.
Generative media 2024 — 2026
Real2SAM2Real: Generative 3D Caches as Complementary Context for Video Diffusion
Manuscript
Control the camera and the objects in video generation with a generative 3D cache
Large-Scale Light Field Synthesis from Videos Enables Geometrically Consistent Bokeh Editing
ECCV 2026
Render geometrically consistent bokeh on scenes that defeat other methods — transparent surfaces, thin structures, cluttered depth
Parametric Shadow Control for Portrait Generation in Text-to-Image Diffusion Models
ICCV 2025
Add a shadow knob to AI portraits, without heavy compute
Flash-Split: 2D Reflection Removal with Flash Cues and Latent Separation
CVPR 2025
Separate reflection from transmission with flash cues and a diffusion prior
Repurposing Pre-trained Video Diffusion Models for Event-based Video Interpolation
CVPR 2025
Turn event streams into smooth video with a video diffusion model
Imaging under hard conditions 2023 — 2025
Temporally Consistent Atmospheric Turbulence Mitigation with Neural Representations
NeurIPS 2024
Low-pass the temporal dimension to strip atmospheric turbulence out of video
Flash-Splat: 3D Reflection Removal with Flash Cues and Gaussian Splats
ECCV 2024
Separate reflections in 3D using flash-induced cues
CodedEvents: Optimal Point-Spread-Function Engineering for 3D-Tracking with Event Cameras
CVPR 2024
Engineer the PSF so event cameras encode depth for free
Snow Removal in Video: A New Dataset and A Novel Method
ICCV 2023
Remove snow from video consistently across frames, learned from real snowfall rather than synthetic
Perceptual image quality 2020 — 2023
Assessor360: Multi-sequence Network for Blind Omnidirectional Image Quality Assessment
NeurIPS'23
Assess 360° image quality the way a viewer actually explores the sphere
PIPAL: A Large-Scale Image Quality Assessment Dataset for Perceptual Image Restoration
ECCV 2020
Score image quality on the artifacts that modern restoration models actually produce
PDFDatasetNTIRE'21 ChallengeNTIRE'22 Full-RefNTIRE'22 No-Ref
Efficient image restoration 2020 — 2022
Chief Vibe Officer
Presides over the Maryland headquarters. Contributes to every paper by lying on the keyboard. Has strong feelings about not being in the acknowledgements.
