What I’ll remember from CVPR 2026 is moving from one poster to the next and ending up in conversations that changed how I think. I spent the week around computational photography, diffusion, image restoration, and generative vision, and I’m heading back with a lot to reflect on.
LeCun's JEPA has evolved into a vision-language model, with 1.6B parameters rivaling the 72B Qwen-VL.
Instead of predicting words directly, the proposed VL-JEPA learns to predict the core "meaning" of a text in an abstract space, ignoring surface-level wording variations.
This method outperforms standard token-based training with 50% fewer parameters. It beats models like CLIP & SigLIP2 on video classification/retrieval tasks and matches larger VLMs on VQA, while using a decoder only when needed to cut decoding ops by nearly 3x.
VL-JEPA: Joint Embedding Predictive Architecture for Vision-language
Paper: https://t.co/rGglBXvKex
Our report: https://t.co/TXEHRquSBr
Make it go viral! REPOST it
BIG BREAKING 🇧🇩: Radical Islamists, along with the Bangladeshi Army & police, are launching deadly attacks on Hindus in the Chittagong district in BANGLADESH !
The Bangladeshi Army is seen destroying CCTV cameras installed in Hindu homes, brutally assaulting residents, & taking them to unknown locations. This video serves as proof.
It all started with a social media post. Mohammad Osman Molla, a Radical Islamist from Hazari Lane in Chittagong (known as the “Gold Avenue” of the area), posted on Facebook claiming “ISKCON is a terrorist organization.”
ISKCON, as we know, is a Hindu religious organization that follows the teachings of Lord Krishna. In response to the post, local Hindus- most of whom are jewelers- went to Osman Molla’s shop to confront him. Osman quickly apologized, which seemed to settle the matter (his apology letter is attached in another post of this thread). The Hindus left peacefully.
However, Radical Islamists falsely reported to the Army & police, claiming that Hindus had attacked Muslims. Without verifying, the Army entered the Hazari Lane area & began destroying CCTV cameras outside Hindu homes/shops. Collaborating with the Radical Islamists, they launched lethal assaults on Hindu households and shops. So far, one Hindu life has been lost, with fears that this number may increase.
I am in constant contact with Hindus in the Chittagong district and will continue to post updates in this thread.
(Join me on Telegram👇 as well for the news updates. https://t.co/quAcafbZD8 )
A huge disturbance is going on in Chittagong's Hazari Gull after demonstration against FB post calling ISCKON a terrorist org.
Bangladesh Army, Police and Radicals have started attacking Hindus and their houses. Even CCTV's are being broken to hide the slaughter of hindus
Excited to share that our paper, "MetaVIn: Meteorological and Visual Integration for Atmospheric Turbulence Strength Estimation," is accepted at #WACV2025! Huge thanks to my PhD advisor, Suren Jayasuriya @imaginglyceum, and internship supervisor, Scott McCloskey @scott_mccloskey.
Marc Levoy, the pioneer of computational photography, has been a major inspiration for me. His work on camera technology is truly magical. Thrilled to meet him at #CVPR 2024!
Excited to present my #CVPR2024 work "Turb-Seg-Res" (https://t.co/D4dTO5gfcT) today at the Computational Cameras and Displays workshop (11:15am-12:30) and Friday at the Main conference (5-6:30pm) at Seattle Convention Center | Arch 4A-E.
Attending #CVPR 2024 in Seattle! Excited to connect with old and new friends. Looking forward to the latest advancements and discussions in computer vision. If you're here, let's catch up!
Microsoft launched its own AI-based Image creator. Later it will be integrated into their chat service.
I wonder if Large Language models are providing the fuel to get rapid progress in AI.
https://t.co/96BUczVmWu
@soumithchintala Even though Bard wasn't programmed to assist with coding. It can generate code (sometimes). I gained access to Bird and took a snapshot of its capabilities. Let me know, if you would like me to attempt anything further.
Attention can now be applied in sliding-windows efficiently and flexibly with NA/DiNA, similar to what (dilated) Convolutions do to FC layers. They preserve locality, capture global context, and expand RFs exponentially, all with linear complexity!
PwCode: https://t.co/yL3wjYPZWQ