acceptodds
Under review as a conference paper at ICLR 2027

FIRM: FLOW-BASED IMAGING VIA REGULARIZED MINIMIZATION

Abstract

Flow matching methods for inverse problems typically incorporate measurements through network conditioning or guidance during sampling. Neither approach explicitly applies the forward operator within the learned conditional velocity field. We develop a principled measurement-conditional velocity parameterization that does. For a linear interpolation path, we express the optimal velocity through the posterior mean and show that this mean is the unique minimizer of a variational objective with an explicit data-consistency term. The velocity defined by this minimizer provably transports the source distribution to the measurement-conditioned posterior. This result leads to a forward operator-aware velocity field that alternates data-consistency steps with learned updates, is trained end-to-end, and requires no separate guidance during sampling. Across five imaging tasks, our method achieves leading reconstruction quality with up to fewer network evaluations than competitive flow-based methods. Varying the number of sampling steps also controls the distortion-perception trade-off without retraining.

Primary area:
generative models
Keywords:
flow matching, inverse problems, conditional generative models, posterior sampling, plug-and-play

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_APWiOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes FIRM, a measurement-conditioned flow-matching method for imaging inverse problems. Through a variational characterization of the posterior mean, the authors motivate incorporating explicit data-consistency updates into the conditional velocity field. The resulting architecture uses an unrolled half-quadratic splitting module and is trained end-to-end. Experiments on five image restoration tasks across two datasets demonstrate competitive reconstruction quality and reduced inference cost.

Critical strengths

1.Clear motivation and a coherent connection between theoretical analysis and architectural design.

The paper addresses how measurement information should enter the generative process by connecting posterior-mean estimation, variational optimization, and conditional velocity parameterization. Explicitly applying the measurement operator within each velocity-field evaluation gives the architecture a clear model-based structure. The theoretical characterization provides a reasonable foundation for integrating optimization unrolling with conditional flow matching.

2.Broad experimental evaluation with competitive reconstruction quality and inference efficiency.

The evaluation covers multiple restoration tasks and includes larger test sets, architectural ablations, and mismatch experiments. The capacity-matched CelebA comparisons provide convincing evidence of effectiveness. In the reported CelebA deblurring configuration, FIRM requires 10 network evaluations and 0.24 seconds, compared with 500 evaluations and 13.03 seconds for Flower5-OT, demonstrating a substantial inference benefit.

Critical weaknesses

1.Differences in comparison conditions leave some uncertainty about the source of the improvements.

In the conditional-flow ablation, matching the sampling steps gives FIRM more internal network evaluations, while matching evaluation counts changes the sampling steps and the corresponding distortion–perception operating point. Some AFHQ-Cat models also have greater capacity than the flow baselines. The existing ablations provide positive evidence, but a fuller discussion of these differences would help clarify the contribution and scope of operator embedding.

2.The relationship between sampling steps and the distortion–perception trade-off could be explained more thoroughly.

The paper demonstrates how changing the sampling steps affects reconstruction metrics, but some settings exhibit non-monotonic behavior, including simultaneous deterioration in PSNR and LPIPS. Discussing these observations in relation to conditional-mean estimation, posterior sampling, and numerical approximation would strengthen the experimental analysis.

Questions for authors

1.Could the authors clarify which conclusions are supported by comparisons at matched sampling steps versus matched network evaluation counts? Based on the existing ablations, how should we interpret the respective contributions of explicit operator embedding and additional internal computation?

2.Could the authors further explain how changing the sampling steps affects the reconstruction distribution, and discuss possible reasons for the simultaneous deterioration in PSNR and LPIPS in some settings?

4 comments

We thank the reviewer for the feedback. In response, we have added a capacity-matched baseline without the data-consistency step (W1, Q1), an analysis of sampling steps from N=1N=1 to N=100N=100, and a comparison with the exact posterior using a Gaussian random field (W2, Q2). See our point-by-point response below.

Differences in comparison conditions leave some uncertainty about the source of the improvements. In the conditional-flow ablation, matching the sampling steps gives FIRM more internal network evaluations, while matching evaluation counts changes the sampling steps and the corresponding distortion–perception operating point. Some AFHQ-Cat models also have greater capacity than the flow baselines. The existing ablations provide positive evidence, but a fuller discussion of these differences would help clarify the contribution and scope of operator embedding.

We disentangle these factors through three controlled comparisons, each holding a different condition fixed (CelebA, PSNR / SSIM / LPIPS, better values in bold).

1. Matched sampling steps. At equal NN, FIRM uses more NFEs but achieves higher PSNR and SSIM in all 10 settings and lower LPIPS in 7. On average, FIRM improves PSNR by 0.78 dB and SSIM by 0.010, while reducing LPIPS by 10%.
TaskCond. flow, N=2N=2FIRM, N=2N=2Cond. flow, N=10N=10FIRM, N=10N=10
Box inp.31.03 / 0.941 / 0.031633.25 / 0.960 / 0.022529.58 / 0.934 / 0.028131.85 / 0.953 / 0.0206
Denoise32.89 / 0.917 / 0.039133.67 / 0.932 / 0.034431.49 / 0.897 / 0.039632.26 / 0.914 / 0.0320
Rand. inp.34.54 / 0.955 / 0.018635.03 / 0.963 / 0.018933.53 / 0.948 / 0.014933.98 / 0.956 / 0.0146
SR34.24 / 0.947 / 0.029134.39 / 0.949 / 0.028533.04 / 0.935 / 0.023933.13 / 0.937 / 0.0228
Deblur35.77 / 0.952 / 0.028136.04 / 0.957 / 0.029534.72 / 0.943 / 0.020935.03 / 0.950 / 0.0235
2. Matched computational budget. At equal NFEs (conditional flow at N=10N=10 versus FIRM at N=2N=2), FIRM improves average PSNR by 2.0 dB and SSIM by 0.021. Conditional flow achieves 5% lower LPIPS on average, reflecting the different distortion–perception operating points.
TaskCond. flow, N=10N=10FIRM, N=2N=2
Denoise31.49 / 0.897 / 0.039633.67 / 0.932 / 0.0344
Deblur34.72 / 0.943 / 0.020936.04 / 0.957 / 0.0295
SR33.04 / 0.935 / 0.023934.39 / 0.949 / 0.0285
Rand. inp.33.53 / 0.948 / 0.014935.03 / 0.963 / 0.0189
Box inp.29.58 / 0.934 / 0.028133.25 / 0.960 / 0.0225
3. No-DC control. Same network, parameters, NFEs, and NN as FIRM; only the data-consistency step is replaced by the identity. On average, FIRM improves PSNR by 0.57 dB and SSIM by 0.006 and reduces LPIPS by 8%, and it is better in every metric and setting.
TaskNNFIRMNo-DC
Deblur236.04 / 0.957 / 0.029535.76 / 0.954 / 0.0310
Deblur1035.03 / 0.950 / 0.023534.74 / 0.946 / 0.0243
Rand. inp.235.03 / 0.963 / 0.018934.18 / 0.954 / 0.0209
Rand. inp.1033.98 / 0.956 / 0.014633.13 / 0.948 / 0.0176

Capacity. On CelebA, all comparisons above are capacity-matched (34.5M, the size of the flow baselines' prior). On AFHQ-Cat, FIRM is larger than the flow baselines' prior (31.0M) but smaller than I2SB (552.8M): 59.9M, and 239.4M for deblurring, the hardest setting at 256×256256\times256, where smaller estimators underfit (App. D.1).

2.The relationship between sampling steps and the distortion–perception trade-off could be explained more thoroughly. The paper demonstrates how changing the sampling steps affects reconstruction metrics, but some settings exhibit non-monotonic behavior, including simultaneous deterioration in PSNR and LPIPS. Discussing these observations in relation to conditional-mean estimation, posterior sampling, and numerical approximation would strengthen the experimental analysis.

Numerical explanation of the non-monotonic behavior. The deterioration previously observed at N=25N=25 results from the velocity scaling used during sampling. The velocity (zθK−xt)/(1−t)(z_\theta^K-x_t)/(1-t) is singular at t=1t=1, so training uses the denominator τt=max⁡(1−t,τmin⁡)\tau_t=\max(1-t,\tau_{\min}) (Eq. 21). During sampling, either τt\tau_t or 1−t1-t can be used; the two coincide for N≤10N\leq10. For N>10N>10, using τt\tau_t reduces the magnitude of the final updates, leaving residual source noise and degrading both PSNR and LPIPS. Using 1−t1-t eliminates this deterioration and yields a consistent distortion–perception tradeoff (PSNR / LPIPS):

NNCelebA, deblurringCelebA, SRAFHQ-Cat, deblurringAFHQ-Cat, SR
236.04 / 0.02934.39 / 0.02829.89 / 0.22929.10 / 0.167
435.78 / 0.02833.96 / 0.02629.42 / 0.18828.50 / 0.141
1035.03 / 0.02333.13 / 0.02228.72 / 0.14527.75 / 0.115
2534.19 / 0.01832.50 / 0.02028.16 / 0.11927.25 / 0.104
5033.80 / 0.01632.11 / 0.01927.86 / 0.11427.02 / 0.101
10033.57 / 0.01631.91 / 0.01927.70 / 0.11126.90 / 0.102

As NN increases, PSNR decreases while LPIPS generally improves, with both metrics approaching a plateau at larger NN. This is consistent with a transition from posterior-mean-like reconstructions toward perceptually realistic samples. We will use 1−t1-t for all results with N>10N>10 and clarify this numerical choice in App. F.1.

Posterior fidelity versus distortion–perception tradeoff. Since the exact posterior is unknown for natural images, PSNR and LPIPS alone cannot establish posterior fidelity. We therefore evaluated FIRM on a Gaussian random field prior over 64×6464\times64 images, for which the posterior is available in closed form (random inpainting; App. F.4).

NNDistance to posterior meanW2W_2 to exact posterior ↓
20.0550.829
100.5060.360
250.7910.162

As NN increases, the samples move away from the posterior mean while approaching the exact posterior, with the W2W_2 distance decreasing by approximately fivefold from N=2N=2 to N=25N=25. This provides direct evidence of improved posterior fidelity and complements the distortion–perception trends observed on natural images.

Could the authors clarify which conclusions are supported by comparisons at matched sampling steps versus matched network evaluation counts? Based on the existing ablations, how should we interpret the respective contributions of explicit operator embedding and additional internal computation?

Please see our response to W1. In short:

  • Matched NN: FIRM is the better estimator along the same trajectory, with higher PSNR and SSIM in every setting and lower LPIPS in 7 of 10.
  • Matched NFEs: at equal cost, FIRM has higher PSNR and SSIM on every task; the conditional flow's lower LPIPS reflects its more perceptual operating point (N=10N=10).
  • No-DC control (same network, NFEs, and NN): removing only the data-consistency step removes the gain, so the improvement comes from the explicit operator embedding, not from additional internal computation.

Could the authors further explain how changing the sampling steps affects the reconstruction distribution, and discuss possible reasons for the simultaneous deterioration in PSNR and LPIPS in some settings?

Please see our response to W2. We quantify how increasing NN moves the reconstruction distribution away from the posterior mean and toward the exact posterior, using the distance to the posterior mean and the 2-Wasserstein distance to the exact posterior. We also explain how the velocity scaling used during sampling caused the simultaneous deterioration in PSNR and LPIPS, and show that correcting it eliminates this behavior.

We thank the authors for their detailed and thoughtful response. The new ablation clarifies the contribution of the paper. These additions address our main concerns.

AuthorsOpenReview ↗

Follow up on

Thank you for considering our rebuttal. We’re glad the new ablation clarified our contribution and that the additions addressed your main concerns. We’d be happy to provide any further details that could support a stronger recommendation.

Reviewer_bFUbOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces FIRM, a conditional flow-matching method for linear inverse problems with Gaussian noise that applies the forward operator inside every velocity evaluation. The authors show that the optimal conditional velocity depends on a posterior mean, and that this mean minimizes an objective with an explicit data-consistency term. They approximate it with an unrolled HQS solver: closed-form data-consistency steps alternating with a shared U-Net, trained end-to-end. The experiments show that FIRM outperforms the baselines with about 10 network evaluations, versus 100–500 for PnP-Flow and Flower.

Critical strengths
  1. The paper is well-written, and the idea and theory are easy to follow and understand.
  2. The proposed method is novel enough, and the experiments confirm the effectiveness of the proposed model.
Critical weaknesses

The experiments are not sufficient to support the paper’s central claims, particularly the claim that the learned field samples the measurement-conditioned posterior and the stated motivation from scientific and medical imaging. I recommend the following additions:

  • The only quantitative check of posterior fidelity is the 2D Gaussian-mixture experiment. That is too low-dimensional to tell whether the method samples the posterior faithfully at image scale. I suggest a Gaussian prior on images of at least 64×64 (for example, a stationary Gaussian random field with known covariance). With a linear operator and Gaussian noise, this gives a closed-form posterior. The authors could then compare FIRM and the baselines (e.g., PnP-Flow, Flower, OT-ODE) against the exact posterior mean and covariance, using metrics such as mean error, covariance, per-pixel variance error, or KL divergence.

  • The introduction motivates the work with MRI, CT, and microscopy, but all main experiments are natural-image restoration on CelebA and AFHQ-Cat. I suggest evaluating on at least one real scientific or medical dataset, such as MRI, CT, or microscopy. A CT experiment would be particularly interesting because the Radon transform has no closed-form data-consistency solution.

  • A key practical limitation of learned generative priors is hallucination when test data differ from the training distribution. I suggest applying the trained super-resolution model, and the baselines, to out-of-distribution data such as microscopy images. It is interesting to see how the posterior standard deviation tracks the actual error. Even if the method degrades, such an experiment would clarify its practical scope and add value to the paper.

Questions for authors

How does the same U-Net trained with a direct MSE loss, with no flow matching, compare in PSNR/SSIM? This would show whether flow matching helps at the N=2 operating point.

4 comments

We thank the reviewer for the feedback. In response, we have added a comparison with the exact posterior using a Gaussian random field (W1), experiments on real multi-coil MRI data (W2), out-of-distribution super-resolution experiments with uncertainty quantification (W3), and a comparison with the same U-Net trained using an MSE loss (Q1). See our point-by-point response below.

The only quantitative check of posterior fidelity is the 2D Gaussian-mixture experiment. That is too low-dimensional to tell whether the method samples the posterior faithfully at image scale. I suggest a Gaussian prior on images of at least 64×64 (for example, a stationary Gaussian random field with known covariance). With a linear operator and Gaussian noise, this gives a closed-form posterior. The authors could then compare FIRM and the baselines (e.g., PnP-Flow, Flower, OT-ODE) against the exact posterior mean and covariance, using metrics such as mean error, covariance, per-pixel variance error, or KL divergence.

FIRM achieves the lowest W2W_2 distance to the exact posterior at N=25N=25. We trained FIRM on a stationary Gaussian random field prior on 64×64 images, for which the posterior is available in closed form. We compared FIRM with PnP-Flow, Flower, and OT-ODE, using the exact prior velocity for the baselines. We generated 2048 samples for each of 6 measurements, under random inpainting (70% of pixels missing) and deblurring, with noise level σ=0.05\sigma = 0.05.

Distance to posterior mean / per-pixel variance error ↓ / W2W_2 ↓

MethodNFEsRandom inpaintingDeblurring
PnP-Flow11000.633 / 0.415 / 0.4350.479 / 0.560 / 0.402
Flower1-OT1000.734 / 0.350 / 0.4090.436 / 0.567 / 0.347
OT-ODE120–1803.670 / 1.396 / 1.2561.039 / 0.042 / 0.276
FIRM, N=2N=2100.055 / 0.944 / 0.8290.020 / 0.982 / 0.874
FIRM, N=10N=10500.506 / 0.455 / 0.3600.471 / 0.531 / 0.357
FIRM, N=25N=251250.791 / 0.196 / 0.1620.817 / 0.189 / 0.128

At N=25N=25, FIRM achieves the lowest W2W_2 distance for both operators, demonstrating the closest agreement with the exact posterior among the evaluated methods. These results are included in App. F.4.

The introduction motivates the work with MRI, CT, and microscopy, but all main experiments are natural-image restoration on CelebA and AFHQ-Cat. I suggest evaluating on at least one real scientific or medical dataset, such as MRI, CT, or microscopy. A CT experiment would be particularly interesting because the Radon transform has no closed-form data-consistency solution.

We added experiments on real multi-coil MRI data. We evaluated FIRM on the fastMRI brain dataset using multi-coil k-space measurements (App. F.5). The forward operator A=PFSA = PFS does not admit a closed-form data-consistency update in general, so FIRM uses conjugate gradients to solve the corresponding subproblem.

MethodPSNR ↑SSIM ↑NFEs ↓
Zero-filled AHyA^{\mathsf H} y25.350.792–
TV26.190.810–
L1-wavelet26.760.800–
Complex U-Net (supervised)32.000.9011
DPS29.220.8501000
FIRM, N=2N=232.680.91310

FIRM achieves the highest PSNR and SSIM among the evaluated methods, outperforming the supervised U-Net by 0.68 dB and DPS by 3.46 dB, while requiring only 10 NFEs compared with 1,000 for DPS. These results demonstrate FIRM's applicability to real medical imaging data with a nontrivial forward operator.

A key practical limitation of learned generative priors is hallucination when test data differ from the training distribution.

Robustness to prior shift. We applied the AFHQ-Cat 4× super-resolution model, trained on cats, on dog and wild-animal images (mild distribution shift) and fluorescence-microscopy images (strong distribution shift), without retraining. We compared the per-pixel standard deviation σ^\hat\sigma with the error of the sample mean (100 images per set; App. F.6).

Test dataPSNRSpearman (σ^\hat\sigma, error) ↑
Cat (in distribution)29.420.52
Dog30.520.46
Wild27.210.49
Microscopy32.850.57

The estimated uncertainty remains positively correlated with reconstruction error across all datasets, including under strong distribution shift. This suggests that FIRM's uncertainty estimates retain useful information about reconstruction reliability even on unseen image classes.

Comparison with baselines. We also evaluated PnP-Flow and Flower on the out-of-distribution data using the same AFHQ-Cat prior (PSNR / coverage).

MethodNFEsCatMicroscopy
PnP-Flow150027.92 / 0.3919.31 / 0.19
Flower1-OT50027.08 / 0.3518.62 / 0.12
Flower5-OT250027.09 / 0.1618.59 / 0.06
PnP-Flow5250027.98 / 0.1719.53 / 0.10
FIRM, N=25N=2512529.40 / 0.8231.85 / 0.55

On microscopy, FIRM outperforms all baselines by more than 12 dB in PSNR while requiring substantially fewer NFEs. These results demonstrate improved reconstruction robustness under a substantial shift in the image distribution.

Robustness to operator shift. FIRM also handles changes in the forward operator at test time by incorporating the operator explicitly rather than learning it implicitly. Our additional experiments (Table 6; App. F.2) demonstrate substantially greater robustness to operator changes than conditional flow models.

How does the same U-Net trained with a direct MSE loss, with no flow matching, compare in PSNR/SSIM? This would show whether flow matching helps at the N=2 operating point.

Comparison with direct MSE training. Following the reviewer's suggestion, we trained the same U-Net architecture using a direct MSE loss and compared it with FIRM on CelebA (PSNR / SSIM / LPIPS; App. F.2, Table 19).

TaskMSE U-NetFIRM, N=2N=2FIRM, N=10N=10
Denoise33.83 / 0.934 / 0.03633.67 / 0.932 / 0.03432.26 / 0.914 / 0.032
Deblur36.06 / 0.957 / 0.03036.04 / 0.957 / 0.02935.03 / 0.950 / 0.023
SR34.35 / 0.947 / 0.03134.39 / 0.949 / 0.02833.13 / 0.937 / 0.022
Rand. inp.35.47 / 0.964 / 0.01835.03 / 0.963 / 0.01833.98 / 0.956 / 0.014

FIRM matches the distortion performance of direct MSE training while enabling posterior sampling. At N=2N=2, FIRM achieves PSNR within 0.5 dB of the MSE U-Net across all tasks, indicating that flow matching incurs little distortion penalty. Unlike direct regression, FIRM also enables a controllable distortion–perception tradeoff through NN. At N=10N=10, it reduces LPIPS by 11–29% relative to the MSE U-Net, at the cost of lower PSNR, while producing posterior samples.

Thank you very much for addressing my concerns. I am happy to raise my score; this is a good paper.

AuthorsOpenReview ↗

Follow up on your updated assessment

Thank you for reviewing our response and raising your score. We’re glad we addressed the concerns you raised. If there are other factors that make you hesitant to give a higher rating, we’d appreciate the chance to address them.

Reviewer_oEhfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper introduces a flow matching method for linear inverse problems. The main idea is how the method uses the forward operator A and measurement y. Usually methods feed the model the measurements/degraded images as input and train it to output the original undegraded sample. Those methods implicitly learn how to constrain samples to the degradation operator. Instead FIRM incorporates the data likelihood into the velocity evaluation by running optimizer steps during sampling. They run experiments with denoising, deblurring, super-resolution and inpaining and show improved PSNR and SSIM compared to the training free PnP-Flow and FLOWER at 50x less compute compared to training free methods.

Critical strengths
  1. Well motivated idea. The posterior-mean estimator is an unrolled half-quadratic-splitting scheme. The forward operator is applied in the velocity evaluation K times through a data-consistency step
  2. Simple and easy to implement
  3. good experimental design: established benchmarks, consistent PSNR/SSIM results compared to baselines. Ablation of the naive A^Ty conditioning in the Appendix table 9.
Critical weaknesses
  1. The method good at predicting the posterior mean with small N, but not necessarily at sampling from the posterior.
  2. weak baselines. Closest baselines should be methods trained on the specific inverse problem like S2IB, which has mostly better LPIPS. How was I2SB trained, what is the capacity?
  3. It is unclear why this design should be better than a plain flow only conditional network from an intuitive or theoretical point. The A^T y ablation is not very convincing. The baseline seems very similar in performance.

Minor weaknesses: Presentation: the blue highlights in the tables for the best method are difficult to see The 50x compute save is not a fair comparison against the training free methods.

Questions for authors

N=1 results would be interesting. Shouldn't the naive conditional flow model be the main baseline? Comparisons on all tasks and settings to that would help to understand whether the more complex design is required.

Same question as weakness 3: 3. Why should this design be better than a plain flow only conditional network from an intuitive or theoretical point. The A^T y ablation is not very convincing. The baseline seems very similar in performance.

5 comments

We thank the reviewer for the feedback. In response, we have added an exact-posterior analysis using a Gaussian random field (W1), evaluations of conditional flow across all 5 tasks and comparisons with task-specific baselines (W2, W3, Q2), and results for N=1N=1 (Q1). See our point-by-point response below.

The method good at predicting the posterior mean with small N, but not necessarily at sampling from the posterior.

FIRM approaches the exact posterior as NN increases. We added an experiment showing that FIRM's samples approach the exact posterior as NN increases (using a Gaussian random field prior on 64×64 images in App. F.4).

NNDistance to posterior meanW2W_2 to exact posterior ↓
20.0550.829
100.5060.360
250.7910.162
PnP-Flow10.6330.435
Flower1-OT0.7340.409

Small NN approximates the posterior mean by design. A single Euler step from t=0t=0 yields x0+v(x0,0,y)=E[x1∣y]x_0 + v(x_0, 0, y) = \mathbb{E}[x_1 \mid y], the posterior mean. This is intended: FIRM transitions transitions from posterior mean estimation at small NN toward posterior sampling as NN increases, consistent with the results above.

weak baselines. Closest baselines should be methods trained on the specific inverse problem like S2IB, which has mostly better LPIPS. How was I2SB trained, what is the capacity?

Task-specific baselines. Our comparisons now include I2SB, ISTA-Net+, and a conditional flow-matching model trained separately for each of the five tasks. Results on CelebA (PSNR / SSIM / LPIPS):

MethodDenoiseDeblurSRRand. inp.Box inp.
I2SB–33.25 / 0.924 / 0.01231.92 / 0.891 / 0.01432.80 / 0.938 / 0.01128.72 / 0.803 / 0.042
ISTA-Net+–35.38 / 0.950 / 0.03533.54 / 0.937 / 0.03834.70 / 0.960 / 0.022–
Cond. flow32.89 / 0.917 / 0.03935.77 / 0.952 / 0.02834.24 / 0.947 / 0.02934.54 / 0.955 / 0.01831.03 / 0.941 / 0.031
FIRM (N=2N=2)33.67 / 0.932 / 0.03436.04 / 0.957 / 0.02934.39 / 0.949 / 0.02835.03 / 0.963 / 0.01833.25 / 0.960 / 0.022

FIRM achieves the highest PSNR and SSIM across all five tasks, while I2SB achieves better LPIPS on three tasks.

I2SB was trained with its official implementation on our data and operators for each task, and sampled with 999 NFEs. It has 105.2M parameters on CelebA (ours: 34.5M) and 552.8M on AFHQ-Cat (ours: 59.9M-239.4M).

It is unclear why this design should be better than a plain flow only conditional network from an intuitive or theoretical point. The A^T y ablation is not very convincing. The baseline seems very similar in performance.

Theory and intuition. Proposition 1 establishes that the optimal conditional velocity minimizes Eq. (13). A conditional flow seeks to learn this velocity implicitly from training pairs, whereas FIRM explicitly incorporates the known forward operator. This distinction allows FIRM to perform better overall and adapt to changes in the operator without retraining. We clarify this motivation in Sec. 3.2.

We provide additional results showing that FIRM outperforms the conditional flow at both the same number of sampling steps and the same number of NFEs. Moreover, because FIRM incorporates the forward model explicitly, the operator can be changed at test time, which the conditional flow does not handle (CelebA; PSNR / SSIM / LPIPS).

(1) Same number of sampling steps (N=2N=2).

TaskCond. flowFIRM
Denoise32.89 / 0.917 / 0.039133.67 / 0.932 / 0.0344
Deblur35.77 / 0.952 / 0.028136.04 / 0.957 / 0.0295
SR34.24 / 0.947 / 0.029134.39 / 0.949 / 0.0285
Rand. inp.34.54 / 0.955 / 0.018635.03 / 0.963 / 0.0189
Box inp.31.03 / 0.941 / 0.031633.25 / 0.960 / 0.0225

On average, FIRM improves PSNR by 0.78 dB and SSIM by 0.010, and reduces LPIPS by 9%.

(2) Same number of NFEs (10 NFEs: conditional flow at N=10N=10, FIRM at N=2N=2).

TaskCond. flowFIRM
Denoise31.49 / 0.897 / 0.039633.67 / 0.932 / 0.0344
Deblur34.72 / 0.943 / 0.020936.04 / 0.957 / 0.0295
SR33.04 / 0.935 / 0.023934.39 / 0.949 / 0.0285
Rand. inp.33.53 / 0.948 / 0.014935.03 / 0.963 / 0.0189
Box inp.29.58 / 0.934 / 0.028133.25 / 0.960 / 0.0225

On average, FIRM improves PSNR by 2.0 dB and SSIM by 0.021; the conditional flow has lower LPIPS on three tasks, as expected from its more perceptual operating point (N=10N=10).

(3) Operator change at test time. Each model is trained at one setting and tested at others without retraining.

TaskTest settingCond. flowFIRM
Random inpaintingmask ratio 0.529.76 / 0.900 / 0.03338.26 / 0.978 / 0.008
mask ratio 0.7 (training)34.54 / 0.955 / 0.01835.03 / 0.963 / 0.018
mask ratio 0.824.18 / 0.759 / 0.14832.10 / 0.936 / 0.034
Deblurringblur width σb=0.8\sigma_b = 0.826.32 / 0.805 / 0.14136.70 / 0.959 / 0.033
blur width σb=1.0\sigma_b = 1.0 (training)35.77 / 0.952 / 0.02836.04 / 0.957 / 0.029
blur width σb=1.2\sigma_b = 1.230.79 / 0.902 / 0.12333.66 / 0.920 / 0.038

Note that FIRM is far more robust to a change of operator.

Presentation: the blue highlights in the tables for the best method are difficult to see

We updated to use a darker blue highlight and additionally mark the best result in bold and the second-best result with an underline.

The 50x compute save is not a fair comparison against the training free methods.

We have clarified this tradeoff in the introduction. Unlike training-free methods that reuse a single prior across tasks, FIRM requires task-specific training. It is designed for applications where the same inverse problem is solved repeatedly, allowing the one-time training cost to be amortized over many inferences. Its primary advantage is therefore inference efficiency. This advantage also extends to task-specific methods: I2SB, which likewise requires training for each task, uses 999 NFEs, corresponding to 20–100× more than FIRM.

N=1 results would be interesting.

We added results for FIRM at N=1N=1 (5 NFEs) across all five tasks on CelebA and AFHQ-Cat (PSNR / SSIM / LPIPS).

TaskCelebAAFHQ-Cat
Denoise33.54 / 0.927 / 0.03233.22 / 0.900 / 0.074
Deblur35.86 / 0.953 / 0.02730.12 / 0.813 / 0.268
SR34.45 / 0.948 / 0.02929.30 / 0.815 / 0.187
Rand. inp.34.45 / 0.952 / 0.02134.29 / 0.928 / 0.052
Box inp.33.54 / 0.958 / 0.02529.35 / 0.917 / 0.074

These results show that FIRM achieves strong reconstruction quality with just a single sampling step.

Shouldn't the naive conditional flow model be the main baseline? Comparisons on all tasks and settings to that would help to understand whether the more complex design is required.

We have added conditional flow as a baseline in Table 1, with comparisons across all five CelebA tasks. We also provide additional comparisons at matched sampling steps, matched computational budgets, and under changes to the forward operator, as detailed in our response to W3.

Same question as weakness 3: Why should this design be better than a plain flow only conditional network from an intuitive or theoretical point. The A^T y ablation is not very convincing. The baseline seems very similar in performance.

Please see our response to W3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.