FIRM paper: operator-aware flow matching cuts network evaluations 50x for imaging
prof_kamilov · x · 2026-10-07
A team from WashU and UW-Madison introduces FIRM (Flow-based Imaging via Regularized Minimization), a new method for imaging inverse problems.
- Idea: Instead of network conditioning or sampling-time guidance, FIRM explicitly applies the forward operator inside the learned conditional velocity field, expressing the optimal velocity via the posterior mean and proving it is the minimizer of a variational objective with an explicit data-consistency term. The resulting operator-aware velocity field is trained end-to-end with no separate guidance at sampling time.
- Results: Leading reconstruction quality across five imaging tasks with up to 50x fewer network evaluations than competitive flow-based methods; 0.24 s per image at the best distortion setting (vs. 13.03 s baseline); only 10 NFEs at N=2, K=5.
- Useful knob: Varying the number of sampling steps controls the distortion–perception trade-off (fewer steps favor accuracy, more steps recover perceptual detail) without retraining.
Related event: FIRM Cuts Network Evaluations 50x for Inverse Imaging(3 posts)→
More from Research
- Preprint Brings Generative Model-Controlled Synthesis Beyond DNA Into Chemistry — anshulkundaje · 2026-10-07
- Hidden Chain-of-Thought Extracted from GPT-6 and Claude via Standard API Tool Calling — jiqizhixin · 2026-10-07
- Information-Dense Synthesis cuts molecular discovery experiments from O(d) to O(log d) — anshulkundaje · 2026-10-07
- UCLA finds hybrid-attention layer ordering shapes multilinguality — alternative orders learn 2.5X faster — UCLA · 2026-10-07
- NYU's RGPO uses adaptive rationale scaffolding to fight reward sparsity in RL for reasoning — newyorkuniversity · 2026-10-07
- UCLA aligns cross-lingual representations in decoder-only LLMs via MoE router outputs — UCLA · 2026-10-07