Muhammad Maaz

I am a professional engineer working in Computer Vision research. I like to initiate and participate in design-level discussions.

[Salman Khan @KhanSalmanH]

🚀 Generative video grounding decodes the spatial tube box by box, in sequence. A 100-frame tube costs hundreds of decoding rounds, and attention drifts off the video as it goes.

Parallel Tube Decoding does the whole tube in 2 rounds. 79× faster, better accuracy, 4B params.

1/2

[Muhammad Maaz @mmaaz60]

@peter_richtarik Maybe AI conferences need to allocate some budget to incentivize good reviewers (not just top reviewers), and something more than a free registration.

[Muhammad Maaz @mmaaz60]

@sama Its bad at coding, sorry

[Muhammad Maaz @mmaaz60]

@sama @sama open source model is releasing today?

[Salman Khan @KhanSalmanH]

🚨 Submit to 1st Video LLMs Workshop at #CVPR2025 by April 15!

📌 Non-archival | 💡 Novel work welcome
🎯 Topics: training, data, eval, safety, long-video LLMs
💸 Cash prize challenges & excellent speakers!
🔗 https://t.co/OfIOMseUhX

🧵 Follow: @vidllms

[Muhammad Maaz @mmaaz60]

@muzairkhattak Congratulations, MashAllah, keep growing ✌️

[Muhammad Maaz @mmaaz60]

Knock knock! 🚪 Who's there? 🤔 It's GLaMM! 🌟

Excited to share that our paper, GLaMM: Pixel Grounding Large Multimodal Model, will be presented at CVPR 2024! 🎉 Tomorrow at 10:30 AM, poster #326! 🔥

GitHub: https://t.co/IXnOR72Oxa

Meet the team at Arch4A-E, poster #326! 👩‍💻👨‍💻

[Muhammad Maaz @mmaaz60]

@CVPR Hi Everyone attending @CVPR

While asking for the poster, try asking the poster by giving the author name who submitted the poster, not always the first author name.

Good Luck #CVPR2024

[Muhammad Maaz @mmaaz60]

Hi Everyone attending @CVPR

While asking for the poster, try asking the poster by giving the author name who submitted the poster, not always the first author name.

Good Luck #CVPR2024

[Salman Khan @KhanSalmanH]

⚡How can complimentary strengths of image & video encoders help Video-LMMs? We present new results + diverse instruction data + benchmark!

🔗 VideoGPT+ : https://t.co/EM5ns1uKn0

[merve @mervenoyann]

LLaVA++ based on Phi-3: https://t.co/shFo2aqWmf
LLaVA++ based on Llama 3: https://t.co/wK8ILGhVO9
So many fine-tuned variants in this org: https://t.co/AJaJz9Wqwb

[merve @mervenoyann]

We've been blessed again with new LLaVA-like models based on LLaMA 3 & Phi-3 🤩
Also passes the baklava benchmark 🤝✅

[Muhammad Maaz @mmaaz60]

@KhanSalmanH @FahadShahbazkh3 @hanoonaRasheed @mbzuai Well, LLaVA #LLaMA 3 seems to be following instructions better than LLaVA-NeXT-34B

Check out demo at https://t.co/LFUOQ213zD

[Salman Khan @KhanSalmanH]

📈Exciting updates to our recent effort to extend LLaMA3 and Phi3 for *visual* understanding. Enjoy!
💻Online demo: https://t.co/yHcQ8niYmd
📓 Chat in Google Colab: https://t.co/C4ULZzSSnd
🚀LoRA, fully FT and S2 FT models added! https://t.co/o4AVEn0AYF

[Muhammad Maaz @mmaaz60]

@mervenoyann @KhanSalmanH @hanoonaRasheed @huggingface Hi @mervenoyann , that would be great if we can host it on Hugging Face spaces. Please let me know how we can proceed. My email is [email protected]

[Muhammad Maaz @mmaaz60]

@skalskip92 @KhanSalmanH @abacaj @FahadShahbazkh3 @hanoonaRasheed @mbzuai It’s available now https://t.co/aFknnVM7yn

[Muhammad Maaz @mmaaz60]

@skalskip92 @KhanSalmanH @abacaj @FahadShahbazkh3 @hanoonaRasheed @mbzuai Hi, the Google Colab demo is available now. Check it at https://t.co/aFknnVM7yn

[Muhammad Maaz @mmaaz60]

@skalskip92 @KhanSalmanH @abacaj @FahadShahbazkh3 @hanoonaRasheed @mbzuai Hi, currently the demo is not available however all the codes and models are available. You can run the demo offline in case if you have resources. We are working on arranging resources to host an online demo.

[Muhammad Maaz @mmaaz60]

@baptistejamin @abacaj Some of these findings are summarized at https://t.co/blJCdj3nnG and some more will be added soon

[Muhammad Maaz @mmaaz60]

@baptistejamin @abacaj When it comes to fine-tuning on a downstream task, LLaMA-3 8B I see more stable than Phi-3. However low rank adaptation, I notice Phi-3 is still the winner in most of the cases.

[Muhammad Maaz @mmaaz60]

🌟 Hey everyone! We are excited to announce our new project: “LLaVA++: Enhancing Visual Capabilities with #LLaMA-3 and #Phi-3, now available on GitHub.

GitHub: https://t.co/blJCdj3nnG

Please do not forget to STAR (⭐) our GitHub Repo (LLaVA-pp). Thanks a ton! :)