GeoDiffusion
A Training-Free Framework for Accurate 3D Geometric Conditioning in Image Generation
🎨 IEEE/CVF International Conference on Computer Vision (ICCV) 2025
Overview
Diffusion models generate stunning images — but controlling their 3D geometry usually requires costly retraining. What if you could get precise geometric control without any training at all?
GeoDiffusion is a training-free framework that injects accurate 3D geometric conditioning into image generation. It enables precise control over scene geometry while maintaining high image quality — no fine-tuning, no additional training, just plug and play.
Key Contributions
- 🆓 Training-free — no fine-tuning or retraining needed to add geometric control
- 📐 Accurate 3D conditioning — precise control over the 3D geometry of generated scenes
- 🎨 High image quality — maintains generation fidelity while enforcing geometric constraints
- ⚡ Efficient and accessible — easily integrable into existing diffusion pipelines
Why It Matters
Geometric consistency is critical for applications like autonomous driving simulation and robotics. GeoDiffusion makes precise 3D-aware image generation accessible to everyone — without the cost of retraining large diffusion models.