xiangxiang chu

Head of DreamX @ Alibaba AMAP. Building product-scale foundation AI for spatial intelligence, agents, world models & generative AI. 300M+ daily users.

[xiangxiang chu @cxx1353574]

I lead AMAP-ML at Alibaba AMAP.

We build product-scale AI systems for spatial intelligence, agents, world models and generative AI, connecting research, open source and products used by 300M+ people every day.

I will share papers, code and AI team lessons here.

[xiangxiang chu @cxx1353574]

@HuggingPapers Thanks for sharing our NeurIPS26 paper. We also release the code https://t.co/vAzf5V78Ye

[xiangxiang chu @cxx1353574]

中文解读：NeurIPS26 1/14: Urbex城市级3D主动搜索 这次NeurIPS2... https://t.co/VeumKlPWYv 复制一下，打开【小红书】就能阅读全文。

[xiangxiang chu @cxx1353574]

NeurIPS26 1/14 DreamX Urbex
redefines city-scale 3D grounding as active spatial search — an agentic VLM that queries landmarks, zooms into regions, and renders oblique views before committing a bounding box.

[xiangxiang chu @cxx1353574]

github：https://t.co/vAzf5V78Ye

[xiangxiang chu @cxx1353574]

NeurIPS 2026 放榜，DreamX 中稿 11 篇，其中 1 篇 Spotlight 🎉
从上一届 0 篇到今年 11 篇，也算记录一下团队这一年的成长。
我们不是纯研究团队，始终坚持业务交付与前沿探索并行。过去一年，团队 AI 顶会论文已超 70 篇。
继续做真正有影响力、能落地的工作。中秋快乐 🌕

[xiangxiang chu @cxx1353574]

@TumatiKiran thanks for sharing our work

[xiangxiang chu @cxx1353574]

@realrebelai thanks for sharing our work

[xiangxiang chu @cxx1353574]

@SlipperyGem thanks for sharing our work!

[ModelScope @ModelScope2022]

Say hello to DreamX-Creator 1.0, a 7B native audio-video generator with local 2K refinement. Apache 2.0.
🤖 https://t.co/mXfsOgor7n

🏆 Delivers results comparable to much larger SOTA models like MiniMax H3 on Verse Bench, reaching 0.1351 DeSync after RL and 0.6930 VQ with the 2K Refiner.
🎬 Turn a first frame and prompt into synchronized video, dialogue, action sounds, ambience, weather, crowds, and music. No separate dubbing stage.
🔊 Gated Cross-Modal Attention lets sound and visuals shape each other, while modality-aware RL improves video, audio, and their alignment.
✨ The one-step 2K Refiner tops the compared refinement methods on MUSIQ (0.7073) and MANIQA (0.4382) while preserving motion and audio timing.

[xiangxiang chu @cxx1353574]

@ModelScope2022 thanks for sharing our work!

[xiangxiang chu @cxx1353574]

@gptzone_net thanks for sharing our work

[xiangxiang chu @cxx1353574]

@osamabinadnan1 thanks for sharing our work

[xiangxiang chu @cxx1353574]

@rsensui thanks for sharing our work

[xiangxiang chu @cxx1353574]

@shao__meng thanks for sharing our work

[xiangxiang chu @cxx1353574]

@Alibaba_Wan thanks for sharing our work!

[xiangxiang chu @cxx1353574]

@princedoesai thanks for sharing our work

[xiangxiang chu @cxx1353574]

@braincramps thanks for sharing our work

[xiangxiang chu @cxx1353574]

Technical report:
https://t.co/gBRRnTZ0yv

[xiangxiang chu @cxx1353574]

🎬 DreamX-Creator 1.0 is here.
Video and audio, generated together—not stitched together afterward.
From a first frame + prompt, our 7B model jointly creates sound and motion. 1-step 2K refinement delivers synchronized 2K output.
Project & roadmap
https://t.co/LcfjTad5xh

[xiangxiang chu @cxx1353574]

@M1ND3XPAND3R @wildmindai yes, we try to release a strong model with fewer parameters