Frame-based Equivariant Diffusion Models for 3D Molecular Generation

Recent methods for molecular generation face a trade-off: they either enforce strict equivariance using costly architectures or relax it to gain scalability and flexibility.We propose a frame-based diffusion paradigm that enforces deterministic E(3)-equivariance while decoupling symmetry handling from the backbone. We study Local Frame-based Diffusion (LFD), which constructs node-specific frames, and Global Frame-based Diffusion (GFD), which assigns a shared molecular frame.To enhance expressivity, we tested EdgeDiT, a Diffusion Transformer with edge-aware attention. On QM9, GFD with EdgeDiT achieves a test NLL of $-137.97 \pm 0.00$, the lowest among evaluated baselines, and obtains state-of-the-art atom stability (98.89\%), molecular stability (89.39\%), and validity (96.04\%) within models of comparable scale.GFD converges faster than the baselines evaluated in our study and generates 10,000 molecules in 0.33 s, indicating improved sampling efficiency.These results establish frame-based diffusion as a scalable, flexible, and physically grounded paradigm for state-of-the-art molecular generation.