🚀 Introducing GenLit – Reformulating Single-Image Relighting as Video Generation! We leverage video diffusion models to perform realistic near-field relighting from just a single image—No explicit 3D reconstruction or ray tracing required! No intermediate graphics buffers,

See you in Hawaii🖖🏝️.

RT @xiuyu_l: Sparsity can make your LoRA fine-tuning go brrr 💨. Announcing SparseLoRA (ICML 2025): up to 1.6-1.9x faster LLM fine-tuning (2….

RT @seohong_park: Q-learning is not yet scalable. I wrote a blog post about my thoughts on scalable RL algorithms.….

🧠 Takeaway: Our approach demonstrates the untapped potential of foundation models—they inherently understand enough physics to tackle complex graphics tasks, bridging generative AI with practical applications in AR, product visualization, and creative editing. Kudos to our.

📈 Results? GenLit outperforms prior state-of-the-art methods across synthetic and real benchmarks, including the challenging MIT Multi-Illumination dataset, with significant gains in visual realism (PSNR, LPIPS, SSIM)! 🧵 4/5

💡 Key insight: Isn't the relighting problem essentially light in motion, interacting with the scene? So why not directly model this as video generation?! We achieve the near-field relighting without 3D by fine-tuning an image-to-video model on a tiny but physical synthetic

🎬 GenLit introduces single image near-field relighting for the first time, meaning our added point lights can freely move within or out of the camera frame. Think of a tiny, glowing fairy light like "Tinker Bell" gracefully illuminating objects and casting realistic shadows! No.

🚨 Come see InterDyn at #CVPR2025!. We're showing how video generative models can simulate physical interactions without explicit simulator! 🌍🎬.📌 Poster #173. 🗓️ Saturday, June 14 | 🕥 10:30–12:30. 📍 ExHall D.🎤 Also catch Rick’s spotlight talk at Agents-in-Interaction.

RT @graceluo_: ✨New preprint: Dual-Process Image Generation! We distill *feedback from a VLM* into *feed-forward image generation*, at infe….

RT @sainingxie: Indeed. For text-to-image, @xichen_pan had a great summary supporting this decoupled design philosophy: "Render unto diffus….

So now we can collect robotics data without teleop??!.

RT @ChungMinKim: Excited to introduce PyRoki ("Python Robot Kinematics"): easier IK, trajectory optimization, motion retargeting. with an….

RT @arthurallshire: our new system trains humanoid robots using data from cell phone videos, enabling skills such as climbing stairs and si….

It’s happening, Peter is presenting our IG-LLM at #ICLR2025. Can we consider the decades-long inverse graphics problem as a code generation task? 🤔. Come by poster #073 now to chat with us!

I will be presenting our SGP-Bench with @ItsTheZhen at #ICLR2025 🚀.Sat, 26th, 3:00–5:30 PM Singapore Time.Hall 3 + Hall 2B Poster #569. Can LLMs 'see' images directly via graphics code?!🧠🖼️ Come by our poster and let's chat!.

RT @RickyTQChen: This ICLR is the best conference ever. Attendees are extremely friendly and cuddly. What do you mean this is the wrong….

Kudos to the amazing team with @junyi42(co-lead), @QianqianWang5, @yufei_ye, @pengcheng_147, @Michael_J_Black, @trevordarrell, and @akanazawa🖖.

Given two images I_i and I_j, instead of independently reconstructing each geometry like DUSt3R or MonST3R, one only needs two properly defined point maps—essentially reconstructing I_j’s 3D moment in I_i’s coordinates—then it yields "Simultanous 4D reconstruction and Tracking in.

At our core, we wonder: What is the minimal representation required to fully model the 4D world? Could it be done in a unified framework? It turns out to be remarkably simple! 🧵3/4.