🌟🔥 Are We Ready for RL in Text-to-3D Generation? A Progressive Investigation

Official repository for the paper "Are We Ready for RL in Text-to-3D Generation? A Progressive Investigation".

[📖 Paper] [🤗 one-step Model]

💥 News

• [2026.02.28] We release the training code! ! 🔥
• [2026.02.21] AR3D-R1 has been accepted by CVPR 2026 Main ! 🔥
• [2025.12.15] AR3D-R1 #3 paper of the day in HuggingFace Daily Papers ! 🔥
• [2025.12.11] We release the checkpoint of one-step AR3D-R1 and the inference code! 🔥
• [2025.12.11] We release the arxiv paper. 🔥

💪 Get Started

Installation

Please set up the Python environment by:

conda env create -f environment.yml
conda activate environment_name

pip install -r requirements.txt

My environment setup is mainly based on ShapeLLM-Omni. If you only need inference, installing this repository is sufficient.

Prepare Reward Model Checkpoints

Please download the reward model you need for training.

cd gen3d-r1/reward_weight

• Download HPS checkpoint from this link by

wget https://huggingface.co/xswu/HPSv2/resolve/main/HPS_v2.1_compressed.pt

• Download Unified checkpoint from this link by

huggingface-cli download CodeGoat24/UnifiedReward-2.0-qwen-7b --repo-type model --local-dir UnifiedReward-2.0-qwen-7b

🚀 Training

cd gen3d-r1/src
bash scripts/run_grpo_3d.sh

Notes:

• Parameters:
  • reward_funcs: The options are hps, unified. You can choose whatever composition you need for training. Make sure to substitute the correct checkpoint path and config path in the run_grpo_3d.sh

💫 Inference

You can download the checkpoint from here

python demo.py

💫 Evaluation with Metrics

We provide an evaluation script that supports both inference and metrics calculation:

python eval.py

Configuration:

• Modify the model_path in eval.py to point to your downloaded checkpoint
• The script is compatible with the inference pipeline and adds comprehensive metrics evaluation
• Supports batch evaluation on test datasets with automatic metric computation

👀 Qualitative result

🗒️ TODO

• Release complete two-step training & evaluation code

• Release one-step training code

🧠 Related Work

• [Image Generation CoT] Can We Generate Images with CoT? Let's Verify and Reinforce Image Generation Step by Step?
• [T2I-R1] T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT
• [ShapeLLM-Omni] ShapeLLM-Omni: A Native Multimodal LLM for 3D Generation and Understanding
• [Trellis] Structured 3D Latents for Scalable and Versatile 3D Generation

⭐Citation

If you find AR3D-R1 useful for your research or projects, we would greatly appreciate it if you could cite the following paper:

@article{tang2025we,
  title={Are We Ready for RL in Text-to-3D Generation? A Progressive Investigation},
  author={Tang, Yiwen and Guo, Zoey and Zhu, Kaixin and Zhang, Ray and Chen, Qizhi and Jiang, Dongzhi and Liu, Junli and Zeng, Bohan and Song, Haoming and Qu, Delin and others},
  journal={arXiv preprint arXiv:2512.10949},
  year={2025}
}