Xiangyue Liu

PhD @HKUST | Multimodal Pretraining, Model Architecture, Large Multimodal Foundation Models

[Xiangyue Liu @star_chenxi]

Fully open-sourced! 🛠️

We made the codebase incredibly clean and minimal so anyone can reproduce it in minutes. Ready for the community to play with! 😄

Might be of interest to:
@_akhaliq @arankomatsuzaki @omarsar0 @huggingface @paperswithcode @fly51fly

[Xiangyue Liu @star_chenxi]

Adding image generation makes your LLM dumber (MoE/MoT collapse)📉. We fixed it!
🪨Rosetta (HKUST 🤝 Tencent Hunyuan Foundation Model Team) eliminates gradient conflicts to keep your LLM smart.
⚡️Strictly ZERO extra VRAM overhead.
💻Single-GPU pretraining reproducible in 90 mins!
https://t.co/lPOgDhzWVp
https://t.co/At0KBi98iR

[Xiangyue Liu @star_chenxi]

See the collapse for yourself.
👀In just 100 training steps, MoE/MoT's ARC scores plummet from ~68 ➡️ ~30. Rosetta stays rock solid.
🚀Try our 1-command demo (~90 mins on 1 or 8x H20):
https://t.co/wenkhfHrgb

[Xiangyue Liu @star_chenxi]

Thrilled to share our paper, Symbiotic-MoE, has been accepted to ECCV 2026! 🎉🎉🎉

❓ How do we unify Multimodal Understanding and Generation in native foundation models without catastrophic forgetting?

💡 We propose Symbiotic-MoE, a native multimodal Mixture-of-Experts architecture. By leveraging Shared Experts as a global semantic bridge, we achieve symbiotic co-evolution and mutual enhancement between text/vision understanding and image generation—with zero extra parameter overhead! 📈

📄 Paper: https://t.co/ECDP178x3F

This work was developed during my internship at @TencentHunyuan. Deeply grateful to the amazing team and colleagues there for the inspiring discussions and support! See you in Malmö, Sweden this Sept.! 🇸🇪🥳

#ECCV2026 #ComputerVision #DeepLearning #MoE #GenerativeAI #UnifiedModel #LMM #ImageGeneration

[Xiangyue Liu @star_chenxi]

#3DV2025 #GaussianAvatarEditor 🥳🥳🥳 Thrilled to share our 3DV paper, an innovative framework for text-driven editing of animatable Gaussian head avatars that can be fully controlled in expression, pose, and viewpoint. Code has released! 🎉🎉🎉
https://t.co/7uvsMfa9hq

[AK @_akhaliq]

discuss: https://t.co/CWHOwexMTK

[AK @_akhaliq]

GaussianAvatar-Editor

Photorealistic Animatable Gaussian Head Avatar Editor

[Xiangyue Liu @star_chenxi]

@ChuanWen15 Thanks！shixiong！

[Xiangyue Liu @star_chenxi]

#CVRP2024 #NeRF2NeRF 🥳🥳🥳Thrilled to share our high score 554 CVPR paper "GenN2N: Generative NeRF2NeRF Translation"🎉🎉🎉, a unified framework for various NeRF2NeRF tasks like NeRF editing, colorization, super-resolution, inpainting. Code coming soon! 🔗https://t.co/PElkYKhJJn

[MrNeRF @janusch_patas]

Code release: "PhysGaussian: Physics-Integrated 3D Gaussians for Generative Dynamics"
https://t.co/0w0zSxf6Jy

[Chen Change Loy @ccloy]

📽️📽️ The code of Rerender A Video is now available at https://t.co/7FnktKlsn6

#SIGGRAPHAsia2023 #SIGGRAPHAsia

[Shuai Yang @ShuaiYang1991]

Excited to share 𝐑𝐞𝐫𝐞𝐧𝐝𝐞𝐫-𝐀-𝐕𝐢𝐝𝐞𝐨, accepted by #SIGGRAPH_Asia 2023, for generating fantastic and coherent videos with #Stable_Diffusion and #Ebsynth. Code is released and try it at: https://t.co/JveqAWH6M6

[Simon Lee @simonxxoo]

WIP

#b3d

[Wenzel Jakob {deprecation notice} @wenzeljakob]

We can also jointly optimize geometry and appearance. In the example below, we determine the albedo and roughness of a Disney BSDF (shown under new viewing/illumination conditions). (6/8)

[Swiss Robotics @swissrobotics]

Two job opportunities at the Swiss government in the field of robotics:
1.) Job with focus on countering mini drones: https://t.co/lqqCstESps
2.) Job with focus on unmanned ground vehicles (MedEvac, Logistics, Search & Rescue, ...): https://t.co/CHCwTPQxSW

#jobs #swissrobotics

[AK @_akhaliq]

Instant Neural Graphics Primitives with a Multiresolution Hash Encoding
paper: https://t.co/XeFvKb6R7w
project page: https://t.co/RP0I0I9VKd
github: https://t.co/IDj7S7ilVu

[70s Sci-Fi Art @70sscifi]

gonzo-sci-fi:

Moebius

[Rana Hanocka @RanaHanocka]

Amazing opportunity to take a deep dive into geometry processing research — no prior experience is required. Apply and help spread the word! 📐

[Sanja Fidler @FidlerSanja]

@NVIDIAAI Toronto AI Lab (https://t.co/RpusM4YOdk) is looking for motivated interns (undergrad./MSc/PhD) to join our group! We are working on 3D/4D vision, simulation for robotics/AV, animation, generative modeling, content creation, and more. Apply here: https://t.co/ZQnINBB0tu

[Oxford Comp Sci @CompSciOxford]

We are hiring 📢 @CompSciOxford and @BalliolOxford are seeking a Fellow and Tutor in Database System. Details: https://t.co/YpbdeOaQxE Closing date 10 January. #computerscience #databasesystems #fellow #tutor