Bernoulli Flow Models hit FID 9.22 in 16 steps on LSUN Churches — no distillation needed
kastnerkyle · x · 2026-10-12
A new arXiv paper proposes Bernoulli Flow Models (BFM), tackling the quality collapse of binary diffusion models at low inference step counts (NFEs).
- Problem: existing binary diffusion models use discrete one-step Markov chains and approximate multi-step likelihood with single-step transitions when cross-step sampling, severely degrading low-NFE quality.
- Approach: BFM defines a unified continuous global Bernoulli probability flow path and derives analytical closed-form posterior transitions over arbitrary intervals — reducing NFEs means re-evaluating the analytic posterior on a new time grid, not skipping steps, eliminating the train-inference mismatch.
- Results: FID 9.22 in 16 steps on LSUN Churches 256×256 (vs. 204.10), with no distillation or extra training, and strong robustness to aggressive NFE reduction.
More from Research
- Ilya Suggests the Brain May Use Backprop; Critic Calls Metaphor Off by Five Orders of Magnitude — JosephJacks_ · 2026-10-12
- Spotify Uses LLM Explanations for RL Credit Assignment, Beating PPO on Delayed Rewards — _reachsumit · 2026-10-12
- InstantFusion: A Shared Latent Interface Connecting Heterogeneous Image Generators — Shanghai-AI-Laboratory · 2026-10-12
- GraphDecisionBench: Benchmarking System One Models on Graph Tasks — Xianliang Yang · 2026-10-12
- Alibaba's Representation Guidance: 459M-Parameter Pixel Diffusion Transformer Beats Larger Models — alibabagroup · 2026-10-12
- Video-RSI: Video Understanding Agents Recursively Self-Improve Their Harness — Bingjun Luo · 2026-10-12