3. Bayesian Flow Networks is a powerful framework compared to diffusion and flow matching even in non-Euclidean space. 👍
Great experience with Yuxuan Song @kb_syx, Jingjing Gong @JingjingGong_, Prof. Zhou @__Hao_Zhou__ , Prof. Liu (a.k.a. JJ)! 🥰
🎉Excited to announce our crystal generation model CrysBFN is accepted to #ICLR2025 as Spotlight paper! CrysBFN achieves ~100x speedup compared to diffusion based model with even better sampling quality! 😍
Project website: https://t.co/a4x4qlROH7
Paper: https://t.co/ybWaPuou0n
Key Findings:
1. CrysBFN achieves ~100x speedup compared to diffusion based model with even better sampling quality! 😍
2. The guidance of uncertainty/entropy in intermediate generation states matters. 🧐