- Rating
- 4(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes BAM, a new model for solving linear inverse problems. BAM has two key aspects: (i) it is degradation-aware, meaning that the forward operator is given as an input at inference time, so a single network covers many inverse problems; and (ii) it draws samples from the posterior distribution in only a few steps by learning a conditional flow map. Importantly, the underlying architecture (built on RAM) is lightweight (36M parameters), which makes inference cheap and fine-tuning accessible. Numerical results on several datasets and inverse problems validate the method, showing strong perceptual quality with only three network evaluations.
Critical strengths
-
While the building blocks of the method are largely well-known techniques (the RAM backbone and flow maps trained via Lagrangian self-distillation), putting everything together is challenging, and the paper does a good job in this regard. Furthermore, some design choices are smart. The rescaling of the measurement through the stochastic interpolant in (7) supports the large dynamic range associated with the different noise levels of flow matching/diffusion, and its benefit is clearly shown in Appendix D.1 (+4 dB). Similarly, the two auxiliary losses in (12) (LPIPS and contrast) are a simple way to improve perceptual quality without resorting to adversarial training. This overall leans me towards a positive assessment of the paper.
-
Numerical results are strong, particularly with only a few steps: with three NFEs, a BAM variant attains the best LPIPS in 38 of 40 settings and the best CMMD in 36 of 40.
Critical weaknesses
-
Perception vs. distortion. The contribution list is not clear about which aspect (perception or distortion) is improved compared to the baselines. Table 1 shows that RAM is better in PSNR in most settings, whereas BAM is better in LPIPS and CMMD. This is expected, since a posterior sample is not meant to beat the MMSE estimator in PSNR (distortion–perception trade-off), but then the claim that BAM "outperforms [...] in sample quality" (abstract and contributions) should explicitly refer to perceptual quality, and the trade-off should be analyzed (see Question 1). Why it is consequential: this is the central claim of the paper, and as stated, it is only partially supported by the results.
-
Missing comparison with InvFusion. InvFusion (Elata et al., 2025) is very close to BAM: a supervised posterior sampler conditioned on the forward operator. However, it is only discussed in Appendix A.2 and is not compared experimentally. While I understand that the backbone is different (it's a transformer-based architecture, and it incorporates the degradation in a different way), and also the training is different, it would be useful to at least compared them in some example. While I understand this might be challenging for the rebuttal, I would appreciate just a comparison in some inverse problem.
Questions for authors
Clarifications (answers could affect my assessment)
-
Distortion–perception trade-off. If BAM is used as an MMSE estimator, does it recover the performance of RAM? More generally, does BAM allow traversing the distortion–perception trade-off, as observed in other flow-map approaches to inverse problems (e.g., [9])? Table 10 partially answers this (the mean of 64 draws outperforms RAM* in four of five problems), but only on FFHQ at a single noise level, and at 64 times the cost. A more complete answer would make the paper stronger, because then RAM is just a particular case of BAM. It would also address Critical Weakness 1.
-
Role of the noise level . The paper states that is not supplied to the network, so that acts "partially blindly", because " is easily estimated from ". However, (i) the sampler in (13) requires explicitly, and (ii) since is a deterministic function of during training, , and the sampling schedule is chosen so that the final step matches , the network effectively receives through . Could the authors clarify this? Moreover, the claim that is easily estimated is not justified by a reference or an experiment, and it is not obvious for measurements such as compressed sensing. Could you report results with an estimated vs. the true ?
-
Köhler and blind experiments. The paper claims that BAM outperforms RAM and Carbajal et al. on the Köhler dataset, but this is shown only qualitatively (Figures 21 and 22). Appendix E.4 states that no ground truth is available in this setting, yet Figure 22 shows a ground-truth crop, and the Köhler benchmark does provide ground-truth images. Could you report PSNR/LPIPS on this benchmark, as is standard? Which is supplied to the sampler for these real images, where the noise level is unknown, and how sensitive is BAM to errors in the estimated operator and in ? Also, this experiment is not really "blind"; I think it is more of a "robustness" experiment, because the model still relies on an external estimator of the degradation (the KPN of Carbajal et al.), as the authors acknowledge in the limitations. Can the authors clarify this?
-
Training cost. While the model is lightweight, training requires 170 hours on 8 H100 GPUs (~1,360 GPU-hours), which somewhat contrasts with the claim that BAM lowers the economic and environmental cost of training imaging models. Why does a lightweight network need a full node for training? Is this because of the selected batch size, or the cost of the JVP in the LSD loss? Also, why is the model initialized from the public RAM checkpoint? Is this initialization necessary, and how does BAM perform when trained from scratch? Otherwise, the cost of training RAM should also be taken into account.
-
Fine-tuning. In Figure 17, BAM* seems to "smooth" the details. This is consistent with Table 6, where fine-tuning on FFHQ improves PSNR but worsens FID (e.g., super-resolution at : 45.5 → 56.5; deblurring: 40.6 → 47.3). Is fine-tuning pushing the model closer to RAM? Do you have any explanation for why this happens?
Suggestions
-
Scope. The method only works on linear inverse problems with additive Gaussian noise, which the authors acknowledge in the limitations. However, I think the "foundation model" and "Anything" framing oversells what is shown: the non-linear JPEG experiment requires fine-tuning (zero-shot BAM obtains 16.1 dB in Table 9), and the noise levels are restricted to . I suggest adjusting the framing accordingly.
-
Architecture. The modification of the RAM architecture (separate measurement and image-state branches fused via (16), Appendix B) is an important aspect of the paper and a contribution in itself. I suggest briefly describing it in the main paper.
-
Related works. I suggest discussing the following closely related papers:
- Multitask learning with stochastic interpolants [1], which also trains a single interpolant-based model to handle multiple tasks, as BAM does across operators.
- The SHARP family [2, 3], which uses restoration networks trained across degradations as priors for inverse problems.
- Other approaches to posterior sampling for linear inverse problems, namely energy models learned via anisotropic diffusion [4] and exact posterior score estimation [8].
- In the paragraph "Conditional diffusion and bridge models sample the posterior": PMRF [5], which is directly related to the perception–distortion discussion in Critical Weakness 1; stochastic interpolants with data-dependent couplings [6], which perform conditional generation from the degraded measurement; InDI [7], which restores images in a few steps by directly iterating from the degraded image; and Palette [10], one of the first conditional diffusion models for image restoration.
I think some of these works, together with InvFusion (currently in Appendix A.2), deserve to be discussed in the main paper, at least in a few sentences in the introduction. The background could be compressed to make room (for instance, presenting both diffusion and flow matching feels a bit redundant).
[1] Negrel, H., Coeurdoux, F., Albergo, M., & Vanden-Eijnden, E. (2025). Multitask Learning with Stochastic Interpolants. NeurIPS 2025
[2] Hu, Y., Delbracio, M., Milanfar, P., & Kamilov, U. (2024, May). A restoration network as an implicit prior. ICLR 2024
[3] Hu, Yuyang, Albert Peng, Weijie Gan, Peyman Milanfar, Mauricio Delbracio, and Ulugbek S. Kamilov. "Stochastic deep restoration priors for imaging inverse problems." ICML 2025
[4] Zilberstein, Nicolas, Santiago Segarra, Eero Simoncelli, and Florentin Guth. "Learning Normalized Energy Models for Linear Inverse Problems." ICML 2026
[5] Ohayon, G., Michaeli, T., & Elad, M. (2025, May). Posterior-mean rectified flow: Towards minimum MSE photo-realistic image restoration. ICLR 2025.
[6] Albergo, M. S., Goldstein, M., Boffi, N. M., Ranganath, R., & Vanden-Eijnden, E. (2024). Stochastic interpolants with data-dependent couplings. ICML 2024.
[7] Delbracio, M., & Milanfar, P. (2023). Inversion by direct iteration: An alternative to denoising diffusion for image restoration. TMLR
[8] Mammadov, Abbas, Ozgur Kara, Kaan Oktay, Iskander Azangulov, Adil Kaan Akan, Hyungjin Chung, James Matthew Rehg, and Yee Whye Teh. "Exact Posterior Score Estimation for Solving Linear Inverse Problems." NeurIPS 2026 (posted on arXiv in June)
[9] Zilberstein, Nicolas, Morteza Mardani, and Santiago Segarra. "Flow Map Denoisers: Traversing the Distortion-Perception Plane for Inverse Problems." NeurIPS 2026 (posted on arXiv in June)
[10] Saharia, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., & Norouzi, M. (2022). Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings.
12 comments
Question 2 (Critical Weaknesses): Evaluate InVfusion on at least one problem/dataset
Missing comparison with InvFusion. InvFusion (Elata et al., 2025) is very close to BAM: a supervised posterior sampler conditioned on the forward operator. However, it is only discussed in Appendix A.2 and is not compared experimentally. While I understand that the backbone is different (it’s a transformer-based architecture, and it incorporates the degradation in a different way), and also the training is different, it would be useful to at least compare them in some example. While I understand this might be challenging for the rebuttal, I would appreciate just a comparison in some inverse problem.
Response
We thank the reviewer for the suggestion. We added an experimental comparison with InvFusion on the FFHQ super-resolution problem in Table 1 ().
Setup. No InvFusion checkpoint is available for this setting. We therefore trained InvFusion with the official code and its FFHQ HDiT architecture, run at , using the bicubic operator and noise level of Table 1. Training took about 45 h on 2 H100 GPUs. We evaluate with the official sampler: 32 Heun steps, i.e. 63 network evaluations. The test protocol is that of Table 1: the same 64 test images, observations and metric implementations.
FFHQ restoration results on 64 images at measurement noise level (). The Gaussian blur has standard deviation (). The ground truth scores 71.71 on MUSIQ and 4.04 on Q-Align.
| Task | Method | NFEs | PSNR (dB) ↑ | LPIPS ↓ | CMMD ↓ | FID ↓ | MUSIQ ↑ | Q-Align ↑ |
|---|---|---|---|---|---|---|---|---|
| 4xSR | InvFusion | 63 | 28.85 | 0.299 | 0.12 | 52.3 | 70.62 | 3.97 |
| BAM | 3 | 28.84 | 0.272 | 0.19 | 50.4 | 68.65 | 3.61 | |
| BAM-FT | 3 | 29.37 | 0.268 | 0.19 | 62.4 | 63.68 | 3.30 | |
| Gaussian deb | InvFusion | 63 | 29.15 | 0.2722 | 0.0637 | 42.18 | 73.34 | 4.166 |
| BAM | 3 | 29.63 | 0.2504 | 0.0987 | 44.98 | 68.78 | 3.661 | |
| BAM-FT | 3 | 30.16 | 0.2446 | 0.0825 | 51.61 | 65.64 | 3 |
Computational cost. We measured both models on the same NVIDIA H100 GPU, restoring one image at a time. Each figure is the median over 16 images, after warm-up.
| BAM | InvFusion | |
|---|---|---|
| Parameters | 35.7M | 170.4M () |
| Network evaluations | 3 | 63 |
| Time per image | 0.62 s | 11.4 s () |
| Peak GPU memory | 1.13 GiB (fp32) | 0.84 GiB (fp16) |
| Problems covered by one model | all problems and noise levels of the paper | one operator and one noise level |
Discussion.
- Accuracy and perceptual quality. BAM matches InvFusion’s PSNR (28.84 vs 28.85 dB) and is better in LPIPS (0.272 vs 0.299) and FID (50.4 vs 52.3). BAM-FT improves the PSNR to 29.37 dB.
- Distributional and no-reference metrics. InvFusion is better on CMMD and on the no-reference metrics MUSIQ and Q-Align. This is consistent with a large diffusion model trained specifically for this single problem.
- Cost. BAM reaches this quality with a smaller network and 21 times fewer network evaluations, which makes it faster. InvFusion’s peak memory at batch size 1 is slightly lower because the official code runs in half precision; both models fit in about 1 GiB.
- Scope. Most importantly, the InvFusion model above serves a single operator at a single noise level, while one BAM network handles every problem and noise level in the paper, and adapts to new operators given at inference time.
We have added this comparison to the revised manuscript and expanded the discussion in Appendix A.2.
Question 1: Distortion–perception trade-off
If BAM is used as an MMSE estimator, does it recover the performance of RAM? More generally, does BAM allow traversing the distortion–perception trade-off, as observed in other flow-map approaches to inverse problems (e.g., [9])? Table 10 partially answers this (the mean of 64 draws outperforms RAM* in four of five problems), but only on FFHQ at a single noise level, and at 64 times the cost. A more complete answer would make the paper stronger, because then RAM is just a particular case of BAM. It would also address Critical Weakness 1.
Response Please see our first comment.
Question 1 (Critical Weaknesses): Distortion–perception
The contribution list is not clear about which aspect (perception or distortion) is improved compared to the baselines. Table 1 shows that RAM is better in PSNR in most settings, whereas BAM is better in LPIPS and CMMD. This is expected, since a posterior sample is not meant to beat the MMSE estimator in PSNR (distortion–perception trade-off), but then the claim that BAM “outperforms […] in sample quality” (abstract and contributions) should explicitly refer to perceptual quality, and the trade-off should be analyzed (see Question 1). Why it is consequential: this is the central claim of the paper, and as stated, it is only partially supported by the results.
Response
We thank the reviewer for this question.
“Sample quality” means perceptual quality. When we write that BAM outperforms RAM in sample quality, we mean perceptual quality, as the visual comparisons in the paper show. The no-reference metrics we added support this (see our response to Reviewer 2). BAM is above RAM and RAM-FT on MUSIQ and Q-Align in all 12 dataset/problem/metric columns of Table 1, for example, 68.78 vs 51.25 MUSIQ and 3.66 vs 2.75 Q-Align on FFHQ deblurring.
We run the same experiment of Table 10 on DIV2K, getting similar conclusions:
BAM as a regressor, at RAM’s cost. BAM’s network predicts a flow map Fθ(xt,yt;t,s). Its diagonal s=t is trained by flow matching, so one evaluation at s=t gives a posterior-mean estimate:
x^=xt−t∗Fθ(xt,yt;t,t), (1 NFE).
This uses the same cost as RAM, and no averaging over 64 draws as in Table 10. We evaluated it on all six settings of Table 1 (FFHQ and DIV2K, σy=0.05), with the Table 1 checkpoint, images, observations and seeds.
PSNR ↑ / LPIPS ↓:
Results:
One network covers the trade-off. The same network gives a near-MMSE estimate in one evaluation (on the diagonal s=t), posterior samples in three evaluations (through the flow-map jumps), and their average (Table 10), which outperforms RAM. In this sense, a RAM-like regressor is a special case of BAM. We highlight that this near-MMSE estimate, because of the extra losses used during training, is still sharper than the true MMSE, as the LPIPS above shows. We will add this experiment to the paper.
We updated both qualitative and quantitative results in Appendix E.6 of the revised manuscript.