Chao Feng

Thinklite-VL is now accepted by #NeurIPS2025 as spotlight🎉 Excited to catch up with old friends and meet new ones in San Diego!

@CzyangChen congrats

Welcome aboard, Ray3! 🎉 Congrats to the team—really proud to be part of it!

Thanks to AK for sharing our paper!🎉 Training a generative critic model to judge responses makes it BETTER at EVERYTHING. Sometimes the best policy comes from good judgment. Your critic model has been hiding its true potential🌟 🚀Introducing LLaVA-Critic-R1, a family of VLMs that serve as both critic and policy in a single model. No policy training. No in-domain task data. Just 40k preference pairs "Is response A or B better?" for Critic RL Training! Result: +5.7% on 26 visual benchmarks including visual understanding, reasoning, even GUI agents. 71.9 7B-Scale SoTA performance on MMMU! Learn to judge, excel at everything🎭 📄 Paper: huggingface.co/papers/2509.00… 💻 Code: github.com/LLaVA-VL/LLaVA…

GPS as a Control Signal for Image Generation

Please join us on poster #369 tomorrow afternoon @CVPR

Work with @CzyangChen , @holynski_ , Alexei A. Efros, and @andrewhowens. Paper: openaccess.thecvf.com/content/CVPR20… Project page: cfeng16.github.io/gps-gen/

Beyond 2D, we can lift a 3D model directly from our GPS-conditioned model by score distillation sampling, which is trained per landmark.

Sharing our #CVPR2025 paper: "GPS as a Control Signal for Image Generation"! 🛰️+✍️ We turn the GPS tag stored in EXIF of photos into a control signal for diffusion models—so they don’t just know what you asked for, but where you want it to look like. Come to see our poster at Friday 13 Jun 10:30 a.m. — 12:30 p.m. (CT) in ExHall D, Poster #250.

Excited to share our CVPR 2025 paper on cross-modal space-time correspondence! We present a method to match pixels across different modalities (RGB-Depth, RGB-Thermal, Photo-Sketch, and cross-style images) — trained entirely using unpaired data and self-supervision. Our approach learns correspondences through contrastive random walks across visual modalities. #CVPR2025 (1/6)

Can AI image detectors keep up with new fakes? Mostly, no. Existing detectors are trained using a handful of models. But there are thousands in the wild! Our work, Community Forensics, uses 4800+ generators to train detectors that generalize to new fakes. #CVPR2025 🧵 (1/5)

Ever wondered how a scene sounds👂 when you interact👋 with it? Introducing our #CVPR2025 work "Hearing Hands: Generating Sounds from Physical Interactions in 3D Scenes" -- we make 3D scene reconstructions audibly interactive! yimingdou.com/hearing_hands/

Hello! If you like pretty images and videos and want a rec for CVPR oral session, you should def go to Image/Video Gen, Friday at 9am: I'll be presenting "Motion Prompting" @RyanBurgert will be presenting "Go with the Flow" and @ChangPasca1650 will be presenting "LookingGlass"

We have released ThinkLite-VL-72B which is trained with only 7.5k sMCTS-selected samples and achieves 79.7 on MathVista🚀 huggingface.co/russwang/Think…

Will VLMs adhere strictly to their learned priors, unable to perform visual reasoning on content never existed on the Internet? We propose ViLP, a benchmark designed to probe the visual-language priors of VLMs by constructing Question-Image-Answer triplets that deliberately deviate from existing data. Check our gallery at vilp-team.github.io & huggingface.co/datasets/ViLP/… To further enhance VLMs’ reliance on visual information, we propose Image-DPO, as elaborated in this thread. w/ @AngCao3 @GunheeLee @jcjohnss @honglaklee

Ever wish YouTube had 3D labels? 🚀Introducing🎥DynPose-100K🎥, an Internet-scale collection of diverse videos annotated with camera pose! Applications include camera-controlled video generation🤩and learned dynamic pose estimation😯 Download: huggingface.co/datasets/nvidi…

🧠💡 What if your 7B model could beat GPT-4o and Qwen2.5-72B—using just 11k training samples? No distillation. No warm-start. Just smart data and reinforcement learning. Inspired by Moravec’s Paradox, we let the model decide what's actually hard. 🚨 New paper: "SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement" We show how ThinkLite-VL-7B achieves SoTA on MathVista—75.1%, surpassing much larger models. 👇 Here’s how we did it: 🔗 arxiv.org/abs/2504.07934 🧠 Code: github.com/si0wang/ThinkL… #AI #VisionLanguageModels #ReinforcementLearning #MachineLearning #LessIsMore

Thank AK for sharing our work💪 We achieve SoTA performance of 75.1 on MathVista using only 11k QA data through RL without any knowledge distillation🚀 Our model and data are fully opensourced at github.com/si0wang/ThinkL… 🤗 Will share details in another tweet later!