FlagOpen releases InsertAny3D and UniVR open-source repositories
FlagOpen (BAAI) published two open-source repositories on GitHub: InsertAny3D, a pipeline for inserting generated 3D objects into Unity scenes, and UniVR, an SFT and RL training framework for Emu3.5 series models.
China context
- Original name
- 北京智源人工智能研究院
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can fork the repositories and adapt the data loaders and reward functions for their own tasks, but must provide Unity integration and model weights.
- For investors
- BAAI's continued open-source releases indicate active development in 3D generation and unified model training, but commercial viability depends on downstream adoption.
Translated from Chinese. Quotes and facts link to the original sources.
InsertAny3D is a server-side pipeline that takes multi-view renders from Unity and an edited image to generate a Gaussian Splatting object and estimate its pose in Unity world coordinates. It supports TRELLIS, SAM3D Objects, and Hunyuan3D as 3D providers. UniVR is an end-to-end SFT and GRPO reinforcement learning framework for Emu3.5 unified generative models, featuring LoRA and full-parameter training, custom vLLM patches for ~2× throughput, and a bring-your-own-task design.
InsertAny3D uses a multi-stage server pipeline: 3D provider generation, Gaussian rendering, GIM matching, pose estimation via depth back-projection, and SAGS extraction to isolate the inserted object. UniVR implements VR-GRPO with format, global, and step-focal rewards, using a Qwen3-VL-30B evaluator and CLIP-feature variance for sub-step selection.
Developers outside China can now reuse BAAI's open-source code for 3D object insertion and unified model training, reducing the cost of building similar pipelines. The repositories provide server-side code and protocols but omit Unity project files and model weights, so adopters must supply their own Unity integration and download weights separately.
For teams building 3D content tools or training unified vision-language models, these repositories offer a starting point that avoids reimplementing complex pipelines. However, the lack of Unity scripts and weights means integration effort remains.
The next verifiable signal is whether the repositories gain external contributors or forks, and whether BAAI releases the Unity-side scripts or model weights. Adoption can be checked via GitHub stars, issues, and downstream projects referencing these repositories.