[AAAI 2026] Debiasing Diffusion Priors via 3D Attention for Consistent Gaussian Splatting

📢 News [2025.12.8] We released the paper [2025.12.8] We released codes. [2025.11.08] 🎉🎉 Our paper "Debiasing Diffusion Priors via 3D Attention for Consistent Gaussian Splatting" has been accepted to AAAI 2026!

---

🎬 Showcase Gallery

Batman	Goku	Pikachu
Iron Man	Mario	Squirrel

---

🔥 What Makes TD-Attn Special?

🚀 Key Innovations

• 🎯 Universal Plugin: Works seamlessly across different 3D generation frameworks
• 🔍 3D-Aware Attention: Eliminates prior view bias through spatial consistency
• 🎨 Controllable Editing: Semantic-specific interventions for precise 3D modifications
• 📐 Multi-View Consistency: Coherent appearance across all viewing angles

---

📋 Abstract

🔬 Click to explore the technical details

We propose a universal framework that eliminates prior view bias in diffusion models to achieve multi-view consistent 3D generation and editing.

Versatile 3D tasks (e.g., generation or editing) distilling Text-to-Image (T2I) diffusion models have attracted significant research interest for not relying on extensive 3D training data. However, T2I models exhibit limitations resulting from prior view bias, which produces conflicting appearances between different views of an object.

This bias causes subject-words to preferentially activate prior view features during cross-attention (CA) computation, regardless of the target view condition. To overcome this limitation, we conduct a comprehensive mathematical analysis to reveal the root cause of the prior view bias in T2I models.

Our Solution - TD-Attn Framework:

🔹 3D-Aware Attention Guidance Module (3D-AAG): Constructs view-consistent 3D attention Gaussians for spatial consistency

🔹 Hierarchical Attention Modulation Module (HAM): Uses semantic guidance trees for precise CA layer modulation

---

🏗️ Architecture Overview

---

⚡ Quick Start

🛠️ Environment Setup

1️⃣ Create Conda Environment

conda create -n TDATTN python=3.9.16 cudatoolkit=11.8
conda activate TDATTN

2️⃣ Install Dependencies

pip install -r requirements.txt

Requirements:

• 🐍 Python 3.9.16
• 🚀 CUDA 11.7+
• 🔥 PyTorch 2.0.1

---

🎮 Usage Examples

🌟 3D Generation

Choose your preferred framework:

# Option 1: LucidDreamer + TD-Attn
cd LucidDreamer-TDATTN
bash run.sh

# Option 2: GCS-BEG + TD-Attn  
cd GCS-BEG-TDATTN
bash run.sh

# Option 3: DreamScene + TD-Attn
cd DreamScene-TDATTN
bash run.sh

✂️ 3D Editing

cd Editsplat-TDATTN
bash run.sh

---

🎯 Key Features

Feature	Description	Benefit
🔄 Multi-View Consistency	Eliminates conflicting appearances across views	Professional-quality 3D assets
🎨 Controllable Editing	Semantic-specific interventions	Precise modifications without artifacts
🔌 Universal Plugin	Compatible with existing frameworks	Easy integration into your workflow
📊 Mathematical Foundation	Comprehensive analysis of view bias	Robust and reliable results

---

🙏 Acknowledgements

This work builds upon incredible contributions from the open-source 3D generation community:

🌟 Core Frameworks

• LucidDreamer - High-quality 3D generation framework
• GCS-BEG - Gaussian splatting techniques
• DreamScene - Scene generation capabilities
• EditSplat - 3D editing framework

🔧 Technical Components

• HRV - View synthesis techniques
• P2P - Prompt-to-prompt editing methods

We thank all researchers and developers who made these open-source frameworks available to the community.

---

📝 Citation

If you find our work useful in your research, please cite our paper:

@misc{jin2025debiasingdiffusionpriors3d,
      title={Debiasing Diffusion Priors via 3D Attention for Consistent Gaussian Splatting}, 
      author={Shilong Jin and Haoran Duan and Litao Hua and Wentao Huang and Yuan Zhou},
      year={2025},
      eprint={2512.07345},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.07345}, 
}