- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper studies learning an unconditional generator of clean signals from corrupted observations, given only a black-box forward simulation of the corruption channel. The authors propose FRRF, which alternates two steps: a convex program reweights a bank of generated samples by matching the empirical observation mean to the bank's forward-simulated measurements under an MMD criterion plus a quadratic penalty, and a small residual field is then trained so the generator moves only by that reweighting. The main theorem gives geometric contraction of the observation discrepancy after T rounds, up to finite-data and computational error, under a condition that the bank contains an affordable reweighting matching the observation mean. Validation is performed on a synthetic 16-dim ring, 50%-masked CIFAR-10 against SCSI and DiffEM, and galaxy restoration.
Critical strengths
S1: The reweighting step is a genuinely useful decomposition. Instead of estimating a density ratio or reconstructing each observation, it reweights the generator's own samples by a convex program that needs only forward simulation of the channel, and the statistical decision and the neural fit become separately analyzable, which is exactly what the theorem tracks. For the simulator-only setting this is a real step past the EM-style baselines (DiffEM, EM-Diffusion) and SCSI's restoration stage, which all require per-observation reconstruction or an evaluable likelihood.
S2: The residual objective is a clean fix for a real failure mode of weighted flow matching: a refitting term moves the generator even at unit weights, and removing it by subtraction introduces negative curvature where the weights are below one. The residual loss removes that term, keeps the curvature positive, and is null at unit weights, so when nothing is requested the generator is exactly unmoved. The transmission experiment tests this well, and the loss is reusable independent of the outer loop.
Critical weaknesses
W1: The paper would benefit from a more organized line of reasoning. The introduction describes the considered problem, but also dives deep into small technical details, while not explicitly stating the contributions until the middle of the 3rd page. Related work ends up being a part of the Introduction itself, which is a very atypical design. The writing is also overly complicated, with the general framework presented before the MMD specialization the algorithm actually uses, making it difficult for the reader to gently enter the topic.
W2: The method is worse than the most direct competitor on the main benchmark, and the theory does not cover the run that produced it. On 50%-masked CIFAR-10, FRRF reaches FID 15.3 at 39.2 A100-hours, while SCSI reaches 13.6 at 31.8 and DiffEM 13.7 at 39.1. In this domain, FID at matched compute is the primary evidence that a method is useful, and there FRRF is above both, with SCSI additionally cheaper; its calibration advantage is real but secondary to FID. The paper's own results compound this: on the anchored path, refresh-only beats refresh-plus-feedback on FID at every recorded checkpoint, so the feedback round, which is the main contribution, does not improve generation under the tested schedules. Meanwhile the recovery theorem assumes the bank contains an affordable reweighting and analyzes the penalized decision, while the CIFAR-10 run uses the constrained variant, which the paper notes defines a different trajectory. This is the main reason my score is at weak rejection.
Questions for authors
Q1: Is there any configuration where the feedback round improves FID at fixed compute, or where FRRF matches SCSI at the same A100 budget?
Q2: Does the coverage condition hold in the reported runs? Is there a computable proxy that indicates a run is in the covered regime?