Jiaqi Wang

Research Scientist at Shanghai AI Laboratory

[DailyPapers @HuggingPapers]

DeepGen 1.0

A lightweight 5B unified multimodal model that outperforms 80B+ giants like HunyuanImage by 28% on WISE and Qwen-Image-Edit by 37% on UniREditBench—proving scale isn't everything

[AK @_akhaliq]

Pref-GRPO

Pairwise Preference Reward-based GRPO for Stable Text-to-Image Reinforcement Learning

[Jiaqi Wang @wjqdev]

🌟Project page: https://t.co/0UGOWyrdGY
📖Paper: https://t.co/VEDcnUI56t

[Jiaqi Wang @wjqdev]

🚀 Pref-GRPO: A pairwise preference-based GRPO method that tackles reward hacking for T2I models

🎨 UniGenBench: A unified benchmark providing comprehensive, fine-grained evaluation for T2I models across 27 dimensions & 20 scenarios

🤗Leaderboard: https://t.co/bTUshFNQeI

[DailyPapers @HuggingPapers]

Explore Pref-GRPO for stable T2I RL & UniGenBench, a comprehensive T2I benchmark, on Hugging Face!

Paper: https://t.co/c28ghJcn7u
Model: https://t.co/3MrTGNAdzm
Leaderboard: https://t.co/5vHIMAZrQB

[Jiaqi Wang @wjqdev]

Thanks for tweeting our work🍻

[AK @_akhaliq]

CODA

Coordinating the Cerebrum and Cerebellum for a Dual-Brain Computer Use Agent with Decoupled Reinforcement Learning

[DailyPapers @HuggingPapers]

SEAgent autonomously learns through experiential feedback, evolving from specialists to generalists. Key components include a World State Model and Curriculum Generator.

Read the paper: https://t.co/jSzI9IHt8P

Try the model: https://t.co/Ftr1KKEK8i

[DailyPapers @HuggingPapers]

Nvidia's got something new

UnifiedReward-Think is here: a multimodal CoT reward model for both visual understanding and generation

https://t.co/k3z5LARosv

[Zhibing Li @ZhibingLi_6626]

🎉 Excited to introduce IDArb! 🎉

Our method can predict plausible and 𝗰𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝘁 geometry and PBR material for 𝗮𝗻𝘆 𝗻𝘂𝗺𝗯𝗲𝗿📷 of input images under 𝘃𝗮𝗿𝘆𝗶𝗻𝗴 𝗶𝗹𝗹𝘂𝗺𝗶𝗻𝗮𝘁𝗶𝗼𝗻𝘀☀️ !

Webpage: https://t.co/GvfyvbEq25

[Jiaqi Wang @wjqdev]

@JunMa_11 Thx! We are preparing the fine-tuning code. Hope to come out within following two weeks.

[Jiaqi Wang @wjqdev]

🚀 We’re excited to announce the release of InternLM-XComposer2.5-OmniLive (IXC2.5-OL), a comprehensive multimodal system designed for long-term streaming video and audio interactions. This fully open-sourced project delivers functionality similar to Gemini 2.0 Live Streaming and OpenAI Her, with standout features including:

🎥 Chat with Streaming Video & Audio
💾 Long-Term Memory for recalling past video experiences
🏆 Competitive Performance across various video and audio perception benchmarks

📄 Paper: https://t.co/oiBfU74lR4
💻 Code: https://t.co/hP881kVnlM
📦 Models: https://t.co/t7hGEuj9Qx

✨ Immerse yourself in multimodal interaction and create your own app today!

#Gemini2 #OpenAI #ChatGPTAdvancedVoice

[Jiaqi Wang @wjqdev]

Thanks so much for tweeting our work!

[AK @_akhaliq]

InternLM-XComposer2.5-OmniLive

A Comprehensive Multimodal System for Long-term Streaming Video and Audio Interactions

[Jiaqi Wang @wjqdev]

Many thanks for your tweeting 👏

[Adina Yakup @AdinaYakup]

InternLM-XComposer-2.5-OmniLive🔥 a specialized generalist multimodal system for streaming video and audio interactions by @intern_lm.

Model: https://t.co/dSqEg2mnK8
✨ Apache 2.0, but a form is required for a commercial license

[Ziwei Liu @liuziwei7]

😻Fine-Grained Visual Attributes for GenAI😻

#NeurIPS2024 🍎FiVA🍊 is a fine-grained visual attributes dataset and a framework that decouples different visual attributes for GenAI

- Project: https://t.co/hhSlc7PFQm
- Code: https://t.co/Ggji0AluDN
- Data: https://t.co/LgRjvcShl1

[Jiaqi Wang @wjqdev]

@GeekTrailAI Thx a lot!

[Jiaqi Wang @wjqdev]

We have released SAM2Long, a training-free enhancement to SAM 2 for long-term video segmentation
🔥 Less error accumulation facing occlusion/reappearance.
⚡️ A training-free memory tree for dynamic segmentation paths, boosting resilience efficiently.
🤯 Significant improvements over SAM2 across 24 head-to-head comparisons on SA-V and LVOS.

Technical Report: https://t.co/jI0WbJDSHr
Github: https://t.co/nxc1WoMVoO
Homepage: https://t.co/zhx7tQuG2R

#AIML #VideoSegmentation #SAM2Long #ComputerVision

[Jiaqi Wang @wjqdev]

@KyeGomezB SAM2Long is a training-free work; the models are identical to SAM2 & SAM2.1. See https://t.co/MWiDHbEaD9 for details.

[Yunxin Li @LyxTg]

🚀Check out VideoVista, our comprehensive video-LMMs evaluation benchmark! We've assessed 33 video Video-LMMs across 27 tasks. Highlights include the latest GPT-4o-Mini, ranked third, and InternLM-XComposer-2.5, the top-performing open-source model.
More: https://t.co/Ey0MIzXIlT

[Yubo Ma @mayubo2333]

Large Vision-Language Models (LVLMs) perform ideally on the understanding of single-page documents like DocVQA, ChartQA.

Here remains an open question🧐: Can LVLMs handle long documents well?
We introduce MMLongBench-Doc!

🌐 Project Page: https://t.co/rdGoATeCW3

🧵(1/7)