Factoid: Brian Cabral leads an imaging group at Meta.
MULTIMODAL AI
-
YOPO Pruning: Efficient Visual Token Processing for MLLMs
By
–
Treat Visual Tokens as Text? But Your MLLM Only Needs Fewer Efforts to See https://
arxiv.org/abs/2410.06169 https://
github.com/ZhangAIPI/YOPO
_MLLM_Pruning/tree/main?tab=readme-ov-file
… -
Meta Sparsh: General-Purpose Vision-Based Tactile Sensing Encoder
By
–
Meta Sparsh is the first general-purpose encoder for vision-based tactile sensing that works across many tactile sensors + many tasks. The family of models was pre-trained on a large dataset of 460K+ tactile images using SSL. To foster new research, we've released code, a
-
Understanding Multimodal LLMs Latest Papers Overview
By
–
I covered the latest multimodal LLM papers lately (but not the evals, as this would be a different beast): https://
magazine.sebastianraschka.com/p/understandin
g-multimodal-llms
… -
AI4Bharat Releases Largest Indian Language Speech Translation Dataset
By
–
AI4Bharat introduces the largest speech translation dataset for Indian languages, featuring 44,400 hours of audio across 13 languages. Learn more: https://
indiaai.gov.in/article/ai4bha
rat-unveils-bhasaanuvaad-speech-translation-dataset-in-13-languages
… #MultilingualAI #BharatAI #IndiaAI #AI4Bharat #BhashaAnuvaad -
Sora’s Screen Transition Mechanism and Prompt Generation
By
–
It seems that Sora has its own mechanism for screen transitions. When looking at the Generate Prompt feature, three additional prompts were generated aside from the input prompt.
-
Robot Data: Continuous and High-Dimensional Compared to Vision and Language
By
–
And compared with vision and language data, robot data is continuous and high dimensional.
-
Robot Data Scale Compared to Language Models: A Significant Gap
By
–
Maybe, but I wouldn’t use the word “soon”: The data in Llama 3 is approx 80,000 times greater than the data collected by π in a year. And compared with vision and language date, robot data is continuous and high dimensional.
-
Sora API leaked and available on Hugging Face
By
–
Sora APIs got leaked and became available via Huggingface. Seems like it is already failing with the 502 error due to high traffic.
— 🚨 AI News | TestingCatalog (@testingcatalog) 26 novembre 2024
"Sora effect"
But the videos are quite impressive in terms of quality. It is worth the wait 👀 https://t.co/f3zOKkSdyA pic.twitter.com/azYZPG5EUOSora APIs got leaked and became available via Huggingface. Seems like it is already failing with the 502 error due to high traffic. "Sora effect" But the videos are quite impressive in terms of quality. It is worth the wait
-
SmolVLM: Fast 2B Parameters Vision-Language Model On-Device
By
–
when the model is so fast and small it falls "outside" of the pareto curve SmolVLM: a fast and high performance 2B params Vision-Language Model for on-device/in-browser use