Research map / Probabilistic and scientific modeling / Variational and generative sampling
FLINT: Coupling Proximal Initialization and Bounded Stochasticity for Flow-Matching Inverse Problems
Junseo Bang, daewon choi, Dong Ju Mun, Se Young Chun
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FLINT, a flow-matching inverse-problem solver that couples proximal initialization of the source noise with bounded stochasticity in the sampling trajectory.
Problem
Generative posterior sampling for imaging inverse problems with flow matching models, where existing solvers split into trajectory-space refinement and source-space (initial noise) optimization that are treated as independent design choices.
Approach
Argues that trajectory-space and source-space methods are fundamentally coupled and combines proximal initialization of the initial noise with bounded stochasticity during sampling, rather than treating the two as independent choices.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add manifold-constrained proximal initialization (Stage 1) instead of random noise (ablation) | FLINT · Gaussian deblurring FFHQ (100) | PSNR: 28.56 -> 29.60 | Improved |
| Add manifold-constrained proximal initialization instead of random noise (ablation) | FLINT · Gaussian deblurring FFHQ (100) | LPIPS: 0.327 -> 0.218 | Improved |
| Add spectrally weighted data consistency (Wiener weighting) (ablation) | FLINT Stage 1 · Gaussian deblurring FFHQ (100) | PSNR: 29.60 -> 30.11 | Improved |
| Add spectrally weighted data consistency (ablation) | FLINT Stage 1 · Gaussian deblurring FFHQ (100) | KID: 0.008 -> 0.005 | Improved |
| Use moderate stochasticity magnitude lambda_s=0.8 vs 0.3 (ablation) | FLINT Stage 2 · Gaussian deblurring FFHQ (100) | PSNR: 29.59 -> 30.11 | Improved |
| Use excessive stochasticity lambda_s=1.8 vs 0.8 (ablation) | FLINT Stage 2 · Gaussian deblurring FFHQ (100) | LPIPS: 0.192 -> 0.328 | Hurt |
| Use bell-shaped instead of front-loaded stochasticity schedule (ablation) | FLINT Stage 2 · Gaussian deblurring FFHQ (100), lambda_s=1.8 | LPIPS: 0.499 -> 0.328 | Improved |
| Use bell-shaped instead of front-loaded schedule at small magnitude (ablation) | FLINT Stage 2 · Gaussian deblurring FFHQ (100), lambda_s=0.3 | PSNR: 29.60 -> 29.59 | No clear effect |
| Perturb proximal initialization within sufficiency regime (C0 up to about 0.4) (ablation) | FlowDPS with SLERP-perturbed inversion · Gaussian deblurring FFHQ (100) | PSNR/LPIPS | No clear effect |
| Increase cumulative stochasticity G during guided sampling (ablation) | FlowDPS · Gaussian deblurring FFHQ (100) | reconstruction quality | Hurt |
| Combine FLINT with 30 total NFEs vs baselines at 50 NFEs | FLINT · FFHQ SR x8, Gaussian deblurring | PSNR, LPIPS | Improved |
| Apply FLINT to Stable Diffusion 1.5 | FLINT · FFHQ SR x4 (200 NFE) | FID: 37.35 -> 22.36 | Improved |
Closest papers by meaning
- Flow Priors for Linear Inverse Problems via Iterative Corrupted Trajectory Matching (NeurIPS 2024 · similarity 0.90)
- Fast samplers for Inverse Problems in Iterative Refinement models (NeurIPS 2024 · similarity 0.90)
- Latent Refinement via Flow Matching for Training-free Linear Inverse Problem Solving (NeurIPS 2025 · similarity 0.89)
- Solving Inverse Problems with FLAIR (NeurIPS 2025 · similarity 0.89)
- Improving Diffusion Posterior Samplers with Lagged Temporal Corrections for Image Restoration (NeurIPS 2026 · similarity 0.88)
- Provably Robust Score-Based Diffusion Posterior Sampling for Plug-and-Play Image Reconstruction (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
