Ajad Chhatkuli

Diffusion models need exact noise level schedules to work. But recently some models have been shown to work without explicit noise conditioning. How? We show that these time-invariant fields implicitly implement a Riemannian gradient flow on some energy landscape. 1/3

some of our work on (not) using CLAUDE.md/AGENTS.md has been widely profiled recently.

@benjaminbardou Interesting how the art style slowly changes to photo-realism on the background

Revolution [a complete circle or turn round a centralpoint, axis] [to roll back, to bring back]

@GoogleDeepMind is there any timeline for code release? would have been great for cvpr rebuttals ;)

We believe this research could have unlimited applications in the real-world. From providing spatial awareness in robotics 🤖, leveling up efficiency in AR 🕹️, and expanding the capabilities in world models 🌐 D4RT’s potential is a necessary step on the path towards AGI. Find out more → goo.gle/introducing-d4…

We're helping AI to see the 3D world in motion as humans do. 🌐 Enter D4RT: a unified model that turns video into 4D representations faster than previous methods - enabling it to understand space and time. This is how it works 🧵

very good video on hyper-connections and the mHC DeepSeek paper by @jbhuang0604 youtube.com/watch?v=jYn_1P…

@jponttuset Maybe because Zurich 3D looks more accurate ;)

Questioning my life's choices

🆕 Separate Deadlines for #CVPR2026 To improve system stability and provide a clearer submission process, we have just introduced 2 new deadlines that are now separate from the Abstract and the Paper Submission deadlines.

🎤 INSAIT at #ICCV2025! On the first conference day, Dr. Jan-Nico Zaech, research scientist at INSAIT, delivered a keynote at the “Foundation Data for Industrial Tech Transfer” workshop. 🤖 Theme was: building more capable robotic AI beyond robot demonstrations, leveraging human videos and other non-robotic data to learn skills and improve robustness for real-world deployment. As the premier venue in computer vision, ICCV is where perception breakthroughs meet robotics - turning pixels and 3D understanding into action. This bridge is essential for taking embodied AI from the lab to the real world demos. 🧠 INSAIT research powering the story: → MotoVLA: learning new VLA skills from human videos. → ReVLA: making VLAs robust to change and new environments. → GaussianVLM: connecting VLMs with Gaussian Splatting scenes for embodied open-world understanding. → Articulate3D: our dataset modeling parts, joints, and interactions, enabling robots to reason about how things move. 🏭 Impact for tech transfer: From assembly and inspection to mobile manipulation, using diverse non-robot data helps models generalize across tasks, scenes, and embodiments, cutting data costs while improving reliability. #ICCV2025 #INSAIT #Robotics #ComputerVision #EmbodiedAI #FoundationModels #VLA #3DVision #RobotLearning

🌏 Join the 6th Annual Nepal AI School (ANAIS 2025)! 10 days of lectures, labs & hackathons + 1 day of hiking 💡 Topics: LLMs, Geometric DL, Agentic AI, Reasoning, CV & more 📅 Dec 29, 2025 – Jan 8, 2026 🕓 Apply by Oct 16, 2025 🔗 anais.naamii.org.np forms.gle/vgVuuEsAmqqQ97…

New video! A quick dive into the recent Hierarchical Reasoning Model (HRM) through the lens of algorithm synthesis. Check it out: youtu.be/RK7lysjz_G0

We are excited to introduce Ultra3D, a novel 3D generation model that produces extremely detailed 3D models. Project Page: buaacyw.github.io/ultra3d/ Arxiv: arxiv.org/abs/2507.17745

@kwangmoo_yi We had a similar idea in this cvpr’24 paper. openaccess.thecvf.com/content/CVPR20… Would be interesting to try it on GS based works.

Preprint of (not) today: Nguyen et al., "Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video" -- arxiv.org/abs/2504.19819 Better surface-based NeRF via neural field modelling of also the camera pose.

⌛ Open Faculty Positions at INSAIT: Apply by January 30, 2025. We are hiring tenure-track and tenured faculty in computer science and AI, with focus areas including computer vision and robotics, natural language processing (NLP), algorithms & theory, machine learning, information security, safe and secure AI, and foundational models. ⭐ World-Class Conditions: INSAIT provides exceptional working conditions in terms of facilities, computational resources, and compensation packages. 🌏 Global Impact: In just over two years, INSAIT has attracted a diverse team of top faculty, postdoctoral researchers, and PhD students from 20+ nationalities. 🏆 Research Excellence: INSAIT researchers have made significant contributions to science published in leading conferences such as CVPR, NeurIPS, ICML, ICLR, ECCV, ICCV, STOC, FOCS, and SODA, with over 50 publications. Join us in advancing AI research and fostering innovation with transformative global impact. ⏳ Application deadline: January 30, 2025 👉 Apply here: insait.ai/faculty

🎓 INSAIT announces the opening of applications for its summer research internship program - SURF 2025! 🌟 🌍 This is the 3rd edition of SURF, which in its past 2 editions already attracted some of the most talented young researchers in AI and computing from 60+ countries. 🏫 📜 The program has been highly successful so far with participants partaking in intense research on some of the hottest themes in AI today, supervised by world-class mentors at INSAIT. ✈️ In addition to a salary of 1500 EUR/month, INSAIT covers transportation to Sofia and back, as well as accommodation. INSAIT also assists with visa and document processing, and it also organizes social trips and fun events during the internship period. 👉 Don’t miss the opportunity to be part of one of the most exciting AI internship programs today. Application deadline: 3 March 2025. 🔗 Details and application link available in the comments. 💬 #SURF2025 #INSAIT #internship

If you like cycle consistency, there is a stricter form of it: G(G(X)) = X that can describe completion functions. Check out our eccv work on self-supervised shape completion today at poster #206. arxiv.org/abs/2409.15939

@songyoupeng @paschalidoud_1 Congratulations Songyou!👏

@paschalidoud_1 thanks so much Despi!!!

Honored to receive the ECVA PhD Award at #ECCV24! This incredible PhD journey wouldn't have been possible without the tremendous support of two of my wonderful supervisors @mapo1 and @AutoVisionGroup! My deepest gratitude to you both❤️

🌟 INSAIT is hiring! We are actively recruiting faculty (tenure-track, tenured) in Computer Vision & Robotics! Successful candidates will benefit from INSAIT’s outstanding conditions and attractive packages. 💡 Apply here: insait.ai/faculty/. More on INSAIT: insait.ai

📣 Google/ETH PhD position available The group of Prof Schindler at ETH offers a fully funded PhD position in collaboration with our team at Google on the topic of 3D computer vision and generative models. For more details (including how to apply): jobs.ethz.ch/job/view/JOPG_…

@chhatkuliA Project Page: pramishp.github.io/iHuman/index.h…

Very proud to have been involved in this project. First ECCV publication from Nepal. Pramish and Anubhav did a great job with clear motivation of having a fast and elegant approach to recovering animatable and renderable human model.