#ICCV2025 Introducing💡LightSwitch💡- A multi-view material-relighting diffusion pipeline that directly and efficiently relights any number of input images to a target lighting & do 3D asset relighting with gaussian splatting! 🧵

Excited to release our new preprint - we introduce Adaptive Patch Transformers (APT), a method to speed up vision transformers by using multiple different patch sizes within the same image!

🌲Introducing our paper A generalizable light transport 3D embedding for Global Illumination https://t.co/qW04w1NdLU. Just as Transformers learn long-range relationships between words or pixels, our new paper shows they can also learn how light interacts and bounces around a 3D

I’ll be presenting our ICCV poster tomorrow from 2:45-4:45 #293! Come see a fast and effective relighting diffusion method that outperforms SOTA inverse renderers for any # of views 🤩

I’m at ICCV! If you’re around and wanna talk about research hmu 🤙

Introducing Kaleido💮 from @AIatMeta — a universal generative neural rendering engine for photorealistic, unified object and scene view synthesis. Kaleido is built on a simple but powerful design philosophy: 3D perception is a form of visual common sense. Following this idea,

Today was my last day interning @RealityLabs, I had an amazing time in Redmond working on generative models for augmented reality with super cool people! I’ll have a 3 day break before coming back into Meta as a visiting researcher lol Onwards to the next adventure 🌟

Introducing: Hyperscape Capture 📷 Last year we showed the world's highest quality Gaussian Splatting, and the first time GS was viewable in VR. Now, capture your own Hyperscapes, directly from your Quest headset in only 5 minutes of walking around. https://t.co/wlHmtRiANy

SGS-1 is akin to nano banana or runway alpha but for 3D - it targets precise 3D creation and editing via CAD. IMO this is very promising in comparison to concurrent 3D generative creation and editing applications that haven’t seen wide adoption (remember DreamFusion?)

💡 Introducing LuxDiT: a diffusion transformer (DiT) that estimates realistic scene lighting from a single image or video. It produces accurate HDR environment maps, addressing a long-standing challenge in computer vision. 🔗Paper: https://t.co/6cW6WlREBl

DINOv3 seems very good at matching objects across environments as well

In VR, users can experience “magical” interactions, such as moving distant virtual objects with the Go-Go technique. How might we similarly extend people’s abilities in the physical world? 🪄 Excited to share Reality Promises, our #UIST2025 paper, led by the amazing @m0k4r1 ✨

Wow, this makes gradio demos so much easier to build 💯

Introducing DINOv3 🦕🦕🦕 A SotA-enabling vision foundation model, trained with pure self-supervised learning (SSL) at scale. High quality dense features, combining unprecedented semantic and geometric scene understanding. Three reasons why this matters…

Thanks @_akhaliq for sharing our work 😊 For more details check out our website https://t.co/CTDjP41d6c Code at

#ICCV2025 Introducing💡LightSwitch💡- A multi-view material-relighting diffusion pipeline that directly and efficiently relights any number of input images to a target lighting & do 3D asset relighting with gaussian splatting! 🧵

Check out our website for more results and visualizations! This work was done under my advisors Fernando & @shubhtuls :) 🔗Website: https://t.co/CTDjP41d6c 📄Paper: https://t.co/v66tyAnqjm 💻Code & Weights: https://t.co/y9mpyYZhNn 🤗HuggingFace:

As a result, LightSwitch can relight hundreds of images of real or synthetic data at up to 2K resolution in minutes or even seconds, beating or matching inverse-rendering and direct baselines that either require 10+ hours per scene or produce inconsistent inaccurate relightings.

This enables 3D asset relighting using a 3DGS optimized on original images by either: (1) adding novel-view synthesized into the input image set before diffusion, or (2) freezing 3DGS positions and optimizing appearance only on relit inputs.

We extend the model to relight an arbitrary number of images and scale with inference-time compute by shuffling latent sets at each denoising step—approximating full relighting attention without attending to all views at once and exhausting memory!

Inferring intrinsic properties provides valuable context for relighting data of diverse material composition, and doing so in a multi-view setting enables high quality consistent relighting.