Kangfu MEI's Homepage

Kangfu MEI 梅康夫

I am a research scientist at Google AI Innovation & Research (AIR). I got my Ph.D. degree at Johns Hopkins University , advised by Prof. Vishal Patel , where I work on multimodal generative models.

I got my M.Phil degree (2019 - 2021) at The Chinese University of Hong Kong, Shenzhen with the major in Computer Science and Engineering. I got my B.S degree (2015 - 2019) at Jiangxi Normal University with the major in Computer Science and Technology.

I previously interned at: Adobe Research / DAMO Academy / Kuaishou / JD.COM

CV / Email / Google Scholar / Github / Photography

Research

I have strong interests in:

A. Image, Video, and 3D generation with diffusion and auto-regressive models.
B. Robot imitation learning from videos for general policy generation.

Representative papers are highlighted.

Updates

[Feb 26, 2025] Our effort of multimodal computational photography is accepeted by CVPR25. 📷
[Jan 22, 2025] The generative field community now wins the second attempt by us in ICLR25. 🧠
[Jan 16, 2025] Defensed my phd thesis. I am a Doctor now !!! 👨‍⚕️
[Jan 15, 2025] I gave a talk (in Chinese) about my recent efficent image generation work at TechBeat 📰

2025
	The Power of Context: How Multimodality Improves Image Super-Resolution Kangfu Mei, Hossein Talebi, Mojtaba Ardakani, Vishal M. Patel, Peyman Milanfar Mauricio Delbracio CVPR 2025 PDF / project page A novel approach that leverages the rich contextual information available in multiple modalities including depth, segmentation, edges, and text prompts to learn a powerful generative prior for SISR within a diffusion model framework
2024
	Field-DiT: Diffusion Transformer on Unified Video, 3D, and Game Field Generation Kangfu Mei, Mo Zhou, Vishal M. Patel, ICLR 2025 project page / OpenReview / Talk / Code A generative filed model that can generate image, video, and 3D data in a single unified network architecture. The unification can boost video generation with 3D prior in multi-task learning, with the 10 times smaller model size than SORA.
	Bigger is not Always Better: Scaling Properties of Latent Diffusion Models Kangfu Mei, Zhengzhong Tu, Mauricio Delbracio, Hossein Talebi, Vishal M. Patel, Peyman Milanfar TMLR 2024 arXiv The first work that throughly investigates the scaling properties of the recent trending latent diffsuion models (e.g., the representative StableDiffusion).
2023
	CoDi: Conditional Diffusion Distillation for Higher-Fidelity and Faster Image Generation Kangfu Mei, Mauricio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M. Patel, Peyman Milanfar CVPR 2024 project page / arXiv Faster conditional diffusion that produces high-quality images with 1-4 sampling steps.
	VIDM: Video Implicit Diffusion Models Kangfu Mei, Vishal M. Patel, AAAI, 2023 Oral Presentation project page / arXiv Video Generation Diffusion Models By Using Implicit Motiion Condition.
2022
	Deep Semantic Statistics Matching (D2SM) Denoising Network Kangfu Mei, Vishal M. Patel, Rui Huang ECCV, 2022 project page / arXiv / poster A New General Plug-and-play Component For Denoising
	Latent Feature-Guided Diffusion Models for Shadow Removal Kangfu Mei, Luis Figueroa, Zhe Lin, Zhihong Ding, Scott Cohen, Vishal M. Patel WACV, 2023 project page / code / demo / arXiv We (together with Adobe) conducted this very early exploration of applying diffusion models on shadow removal. This proposes the first instance-level shadow removal task.
2021
	LTT-GAN: Looking Through Turbulence by Inverting GANs Kangfu Mei, Vishal M. Patel IEEE Journal of Selected Topics in Signal Processing [IF: 7.695] arXiv The first turbulence mitigation algorithm that can clearly recover face images captured in a range of 300 meters long.
	AttaNet: Attention-Augmented Network for Fast and Accurate Scene Parsing Qi Song, Kangfu Mei, Rui Huang AAAI, 2021 code / arXiv Strip Attention Module (SAM) and Attention Fusion Module (AFM) are proposed for enhancing the accuracy of semantic segmentation networks with limited computational complexity.
2019
	Higher-resolution network for image demosaicing and enhancing Kangfu Mei, Juncheng Li, Jiajie Zhang, Haoyu Wu, Jie Li, Rui Huang ICCV AIM RAW to RGB mapping challenge, 2019 code / arXiv For the first time, a neural ISP has outperformed a traditional ISP (like Huawei's mobile ISP) and achieved visual quality comparable to that of a DSLR.
	Progressive Feature Fusion Network for Realistic Image Dehazing Kangfu Mei, Aiwen Jiang, Juncheng Li, Mingwen Wang ACCV, 2019 code / arXiv PFFNet was the first dehazing netowrk that uses fully end-to-end neural network architecture without physical gating unit. More than 100 works (untill 2024) adapot our strategy and show its effectiveness.
2018
	Multi-scale Residual Network for Image Super-resolution Juncheng Li, Faming Fang, Kangfu Mei, Guixu Zhang ECCV, 2018 code / bibtex We built MSRN in 2018. It quickly becomes the foundamental component in all image restoration works and has more than 800 citations untill 2024.