Loren

PhD Student in Deep Learning for Egocentric Vision - University of Zaragoza

[Loren @murloren]

Last week, I was at CVPR presenting our latest work, which was awarded as a ✨highlight✨ paper

@rmcantin

[Loren @murloren]

Very happy to announce, that our last work "DIV-FF: Dynamic Image Video feature fields for egocentric vision" with
@rmcantin
and J.Guerrero, has been accepted at CVPR 2025!! https://t.co/p6UglbApSy

Take a look 👇🧵

[Loren @murloren]

Happy to see my work VJEPA 2.1 on Yann LeCun slides

[Chris Offner @chrisoffner3d]

Here is @ylecun ‘s recent lecture on world models at @ETH: https://t.co/TmSMz9FMzI

[Loren @murloren]

@massiviola01 Yes, but the Convnext-tiny looks to perform better than the ViT-Tiny, which is the reason? maybe its inductive biases

[Loren @murloren]

@massiviola01 Any inshights of what is the performance drop so big in the ViT-Tiny? Is there any intermediate option between the tiny and the small ViTs?

[Loren @murloren]

@TrainUplifted @ylecun @AdrienBardes We also released two distilled versions from the ViT-G (2B params): ViT-B (80M params) and ViT-L (300M params)

[Loren @murloren]

I am very happy to share the result of my internship at FAIR (Meta): V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning with @ylecun @AdrienBardes

Our approach learns dense, spatially coherent features from video while preserving strong global understanding

[Grigory Sapunov @che_shr_cat]

7/ The numbers are brutal. NYUv2 depth RMSE drops from an unusable 0.682 to SOTA 0.307. ADE20K segmentation mIoU jumps +23.4 points. Ego4D anticipation hits 7.71 mAP. PCA features go from noisy static to sharp object boundaries.

[Grigory Sapunov @che_shr_cat]

2/ V-JEPA 2.1 by Mur-Labadia, Muckley, and the FAIR team fixes the global-local representation bottleneck. It unifies image and video representation learning into a single encoder. This is a massive step for embodied AI world models.

[Massimiliano Viola @massiviola01]

Thread on VJEPA 2.1🤟

This DEFINITELY flew under the radar: just a few days ago, @AIatMeta released V-JEPA 2.1, taking a massive step toward closing the gap between image and video domains.

For a long time, image backbones were the only option for solving dense vision tasks. This model disagrees, showing that universal spatial understanding also emerges from large-scale video models!🎥

[Loren @murloren]

@ylecun @AdrienBardes V-JEPA 2.1's new feature maps are worth a look. Genuinely. 👀

- Paper: https://t.co/PN41ZNPrAs
- Code and models: ViT-G (2B), ViT-g (1B), ViT-L (300M) and ViT-B (80M) https://t.co/lX8hZnwH25

[Loren @murloren]

@ylecun @AdrienBardes Global video understanding is possible with excellent dense features!!

- 77.7 % Top-1 Acc on SSv2 (new SOTA, motion-metric dataset)
- Competitive 87.7 % on K400 and 85.5 % on Imagenet-1K
- Semantically coherent, spatially aligned and temporally consistent features

[alphaXiv @askalphaxiv]

Yann LeCun and his team dropped yet another paper!

"V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning"

In this V-JEPA upgrade, they showed that if you make a video model predict every patch, not just the masked ones AND at multiple layers, they are able to turn vague scene understanding into dense + temporal stable features that actually understands "what is where".

This key insight drove improvements in segmentation, depth, anticipation, and even robot planning.

[Pascale Fung @pascalefung]

Introducing VL-JEPA: Vision-Language Joint Embedding Predictive Architecture for streaming, live action recognition, retrieval, VQA, and classification tasks with better performance and higher efficiency than large VLMs.

• VL-JEPA is the first non-generative model that can perform general-domain vision-language tasks in real-time, built on a joint embedding predictive architecture.
• We demonstrate in controlled experiments that VL-JEPA, trained with latent space embedding prediction, outperforms VLMs that rely on data space token prediction.
• We show that VL-JEPA delivers significant efficiency gains over VLMs for online video streaming applications, thanks to its non-autoregressive design and native support for selective decoding.
• We highlight that our VL-JEPA model, with an unified model architecture, can effectively handle a wide range of classification, retrieval, and VQA tasks at the same time.

by @Delong0_0 @MustafaShukor1 @TheoMoutakanni @willyhcchung Jade Lei Yu Tejaswi Kasarla @AllenBolourchi @ylecun @pascalefung

https://t.co/oUnjCaMKVv

[Wes Roth @WesRoth]

Yann LeCun explains that large language models are trained on about 30 trillion words, representing nearly all public internet text.

He says it would take a human over 500,000 years to read that much.

But a 4-year-old child sees just as much visual data in their first few years of life.

This shows how much richer and more complex real-world experience is compared to reading text.

Training on the web is huge but it still doesn’t match what a child learns just by living.