cs.CVSep 30, 2026

BAM! Bayesian Anything Model: a foundation model for generative computational imaging

Authors: Alessio Spagnoletti, Charlesquin Kemajou Mbakam, Jonathan Spence, Andrés Almansa, Marcelo Pereyra

Organizations: Laboratoire MAP5, UMR 8145, Université Paris Cité, CNRS · Heriot-Watt University, School of Mathematical and Computer Sciences & Maxwell Institute for Mathematical Sciences

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Generative models are transforming Bayesian computational imaging, yet the field still lacks physics-aware foundation models. Current practice falls into two camps. Large foundation image models are deployed as plug-and-play priors with zero-shot approximate likelihood guidance, which introduces significant bias and computational cost. Physics-aware generative models avoid this bias, but each is tied to a specific dataset, task and instrument. We introduce BAM (Bayesian Anything Model), a lightweight foundation model for few-step, physics-aware posterior sampling that generalises robustly to unseen data and tasks, zero-shot or with minimal finetuning. BAM upgrades the operator-conditioned Reconstruct Anything Model (RAM) backbone (Terris et al.) into a conditional flow map, so instrument physics is specified at inference time rather than fixed during training. BAM has just 36M parameters and is pre-trained jointly on large image corpora and libraries of forward operators. A single network then draws posterior samples in a few steps, with no likelihood approximation and no guidance weights to tune. Across linear inverse problems on FFHQ, AFHQ, LSUN, DIV2K and the Kohler camera-shake benchmark, BAM outperforms in just 3 steps both specialised models and leading zero-shot methods in sample quality, at a fraction of their computational cost. BAM gives the community an accessible entry point to generative computational imaging, lowers the economic and environmental cost of training imaging models, and opens a new path for research on physics-aware Bayesian computational imaging. Official page: https://bayesian-anything-model.github.io/

Figures & tables

Figure 1: BAM! One small network, many imaging problems, few steps. Each tile pairs an observation (left) with a BAM! posterior sample (right). A single 36M-parameter network, with the forward operator supplied at inference time, covers ×4\times 4 super-resolution (DIV2K, LSUN), Gaussian deblurring (AFHQ), inpainting (FFHQ, LSUN), non-linear JPEG restoration at Q=10Q=10 (FFHQ) and blind motion deblurring (Köhler), all in 33 steps. Settings are given in Section 4 .
Figure 2: Pareto frontier drawn by SOTA methods compared to BAM.
MethodNFEsDIV2KFFHQ
  PSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrowPSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrow
Gaussian deblurring
BAM324.480.350.0634.429.630.250.1045.0
BAM ⋆\star324.440.350.0435.630.160.250.0851.6
RAM125.630.430.4452.630.540.351.1594.7
RAM ⋆\star126.000.420.3447.031.530.341.0792.4
Table 1: DIV2K and FFHQ restoration at σy=0.05\sigma_{y}=0.05 . PSNR (dB; ↑\uparrow ), LPIPS ( ↓\downarrow ), CMMD ( ↓\downarrow ), and FID ( ↓\downarrow ). BAM variants use thre e steps; BAM ⋆\star and RAM ⋆\star denote finetuning. Bold (underline) mark the best (second-best) reported values for each dataset and problem; – denotes an unavailable result.
Figure 3: CT reconstruction and 4×44\times 4 -block uncertainty and residual maps. Left to right: ground truth, reconstruction A†y\mathcal{A}^{\dagger}{\bm{y}} , one BAM sample, posterior mean, standard deviation over 64 draws, and residual ∣GT−mean∣|\mathrm{GT}-\mathrm{mean}| . Yellow boxes and 4×4\times strips show the same lung region in every panel.
Figure 4: DIV2K restorations at σy=0.05\sigma_{y}=0.05 , one example per problem. BAM and BAM ⋆\star use three steps. Strips show 4×4\times zoom. More problems and comparisons in Figure 16 .
Figure 5: FFHQ restorations. Deblurring and super-resolution at σy=0.05\sigma_{y}=0.05 ; JPEG at Q=10Q=10 with pre-compression noise σJPEG=0.01\sigma_{\mathrm{JPEG}}=0.01 . BAM and BAM ⋆\star use three steps. Strips show 4×4\times zoom; – marks unavailable results. More problems and comparisons in Figure 17 .
Figure 6: AFHQ restorations at σy=0.05\sigma_{y}=0.05 : inpainting, demosaicing, compressed sensing. BAM and BAM ⋆\star use three steps. Strips show 4×4\times zoom. For compressed sensing, the observation column shows A†yA^{\dagger}y ; – marks unavailable results. More examples in Figure 15 .
Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: yy vs. yσy_{\sigma} ablation. Per-sample PSNR for conditioning on yσy_{\sigma} and yy on the Gaussian deblurring task. Dashed lines indicate the mean PSNR for each variant, while the shaded region shows the per-sample improvement.
Figure 8: Qualitative yy vs. yσy_{\sigma} ablation on LSDIR Gaussian deblurring. The observation, reconstruction obtained by conditioning BAM on the original measurement yy , reconstruction obtained with noise-aware conditioning yσy_{\sigma} , and ground truth, for Gaussian deblurring with noise level σy=0.05\sigma_{y}=0.05 . Strips below show 4×4\times linear magnification.
Figure 9: Architecture ablation. Smoothed training LPIPS for the BAM architecture and the base RAM design with stacked operatorsWe show the first 60 60\, k training steps. Lower is better.
Figure 10: Qualitative architecture ablation on DIV2K inpainting. The observation, reconstruction obtained with the original stacked-operator RAM design, reconstruction obtained with the BAM architecture, and ground truth. Strips below show 4×4\times linear magnification.
Figure 11: Contrast-loss ablation on FFHQ. Removing the contrast regularizer can produce visibly over-saturated reconstructions. The auxiliary term stabilizes output contrast during heterogeneous multi-dataset, multi-operator training. Strips below show 4×4\times linear magnification.
Figure 12: Adversarial-loss ablation. Smoothed training LPIPS during FFHQ super-resolution fine-tuning with and without a discriminator loss. Lower is better.
MethodNFEsPSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrow
Gaussian deblurring
BAM 1-step126.560.3230.33454.38
BAM 2-step229.680.2490.14942.07
BAM 3-step330.630.2320.08940.63
SR ×4\times 4
BAM 1-step126.520.3440.38658.85
Appendix
Table 2: BAM sampling-step ablation on FFHQ at σy=0.025\sigma_{y}=0.025 . Results over 64 images per problem. PSNR ( ↑\uparrow ), LPIPS ( ↓\downarrow ), CMMD ( ↓\downarrow ), and FID ( ↓\downarrow ). Each step requires one network evaluation. Bold marks the best available value per problem and metric; – denotes an unavailable value.
MethodNFEsPSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrow
Gaussian deblurring
BAM 1-step126.530.3240.32954.55
BAM 2-step228.970.2510.15843.36
BAM 3-step329.630.2500.09944.98
SR ×4\times 4
BAM 1-step126.540.3460.39658.14
Appendix
Table 3: BAM sampling-step ablation on FFHQ at σy=0.05\sigma_{y}=0.05 . Results over 64 images per problem. PSNR ( ↑\uparrow ), LPIPS ( ↓\downarrow ), CMMD ( ↓\downarrow ), and FID ( ↓\downarrow ). Each step requires one network evaluation. Bold marks the best available value per problem and metric; – denotes an unavailable value.
Figure 13: BAM sampling-step comparison on FFHQ at σy=0.025\sigma_{y}=0.025 . For compressed sensing, the observation column shows A†yA^{\dagger}y . Yellow boxes select matching regions, with 4×4\times linear magnification below each image.
Figure 14: BAM sampling-step comparison on FFHQ at σy=0.05\sigma_{y}=0.05 . For compressed sensing, the observation column shows A†yA^{\dagger}y . Yellow boxes select matching regions, with 4×4\times linear magnification below each image.
MethodNFEsσy=0.025\sigma_{y}=0.025σy=0.05\sigma_{y}=0.05
  PSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrowPSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrow
Deblurring
BAM327.410.3360.2333.5226.890.3540.2635.62
BAM ⋆\star326.850.2940.1823.8926.410.3110.1924.21
RAM127.850.4171.1752.1127.460.4441.4160.40
LATINO823.370.4400.4864.8121.480.4860.5071.87
Appendix
Table 4: AFHQ. PSNR (dB; ↑\uparrow ), LPIPS ( ↓\downarrow ), CMMD ( ↓\downarrow ), and FID ( ↓\downarrow ), with 64 images. BAM and BAM ⋆\star use three steps. NFEs denote the number of neural function evaluations. Bold marks the best values, and underlining marks the second-best distinct value for each problem, noise level, and metric.
MethodNFEsσy=0.025\sigma_{y}=0.025σy=0.05\sigma_{y}=0.05
  PSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrowPSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrow
Deblurring
BAM325.010.3340.0433.5124.480.3490.0634.43
BAM ⋆\star324.910.3390.0333.8024.440.3530.0435.59
RAM126.080.4010.3741.3525.630.4270.4452.55
LATINO823.290.4730.2874.6222.000.5140.39100.27
Appendix
Table 5: DIV2K. PSNR (dB; ↑\uparrow ), LPIPS ( ↓\downarrow ), CMMD ( ↓\downarrow ), and FID ( ↓\downarrow ), with 64 images. BAM and BAM ⋆\star use three steps. NFEs denote the number of neural function evaluations. Bold marks the best values, and underlining marks the second-best distinct value for each problem, noise level, and metric.
MethodNFEsσy=0.025\sigma_{y}=0.025σy=0.05\sigma_{y}=0.05
  PSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrowPSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrow
Deblurring
BAM330.630.2320.0940.6329.630.2500.1044.98
BAM ⋆\star331.170.2250.0847.3330.160.2450.0851.61
RAM131.070.3270.9483.3730.540.3511.1594.74
RAM ⋆\star132.310.3160.8582.5831.530.3411.0792.38
Appendix
Table 6: FFHQ. PSNR (dB; ↑\uparrow ), LPIPS ( ↓\downarrow ), CMMD ( ↓\downarrow ), and FID ( ↓\downarrow ), with 64 images. BAM and BAM ⋆\star use three steps. NFEs denote the number of neural function evaluations. Bold marks the best values, and underlining marks the second-best distinct value for each problem, noise level, and metric.
MethodNFEsσy=0.025\sigma_{y}=0.025σy=0.05\sigma_{y}=0.05
  PSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrowPSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrow
Deblurring
BAM325.780.1541.8689.3525.380.1602.1590.52
BAM ⋆\star326.160.1551.0249.4125.790.1601.2648.28
RAM125.970.2131.9277.5325.580.2392.2087.09
RAM ⋆\star127.250.3041.8776.9026.590.3332.1186.54
Appendix
Table 7: LSUN. PSNR (dB; ↑\uparrow ), LPIPS ( ↓\downarrow ), FID ( ↓\downarrow ), and CMMD ( ↓\downarrow ) with 300 test images. BAM and BAM ⋆\star use three steps. NFEs denote the number of neural function evaluations. Bold marks the best values, and underlining marks the second-best distinct value for each problem, noise level, and metric.
Figure 15: AFHQ restorations. One example per problem, repeated at both noise levels. BAM and BAM ⋆\star use three steps. For compressed sensing, the observation column shows A†yA^{\dagger}y . Matched yellow boxes show regions magnified 4×4\times .
Figure 16: DIV2K restorations. One example per problem, repeated at both noise levels. BAM and BAM ⋆\star use three steps. For compressed sensing, the observation column shows A†yA^{\dagger}y . Matched yellow boxes show regions magnified 4×4\times .
Figure 17: FFHQ restorations. One example per problem, repeated at both noise levels. BAM and BAM ⋆\star use three steps. – indicates an unavailable result. For compressed sensing, the observation column shows A†yA^{\dagger}y . Matched yellow boxes show regions magnified 4×4\times .
Figure 18: LSUN restorations. One example per problem, at σy=0.05\sigma_{y}=0.05 . BAM and BAM ⋆\star use three steps; RAM ⋆ denotes dataset-finetuned RAM. – indicates an unavailable result. For compressed sensing, the observation column shows A†yA^{\dagger}y . Matched yellow boxes show regions magnified 4×4\times .
Figure 19: Gaussian sparse-view CT on LIDC-IDRI. The upper panels show the sinogram, the pixelwise standard deviation of 64 BAM draws, and the absolute difference between ground truth and the empirical mean. The lower panels compare filtered backprojection, one BAM draw, the BAM mean, RAM, and the reference slice. Yellow rectangles select the same lung region; the strips immediately below show that region at 4×4\times linear magnification.
Figure 20: Multiscale CT variability and absolute mean error. Each row compares the population standard deviation (left) with ∣GT−mean∣|\mathrm{GT}-\mathrm{mean}| (right), recomputed after block averaging each sample and the ground truth. All maps are in [0,1][0,1] image units. Black denotes zero and brighter values indicate larger variability or absolute error. These display ranges enhance contrast at each scale.
 Known kernelSpatially varying
MethodPSNRLPIPSCMMDFIDPSNRLPIPSCMMDFID
BAM25.810.3290.0540.5623.370.3400.2040.32
RAM27.490.3550.2545.8422.850.3880.5050.80
Appendix
Table 8: DIV2K motion deblurring. PSNR (dB; ↑\uparrow ), LPIPS ( ↓\downarrow ), CMMD ( ↓\downarrow ), and FID ( ↓\downarrow ), over 64 images. BAM uses three reconstruction steps.
Figure 21: Motion deblurring in three settings. Matched yellow boxes show regions magnified 4×4\times .
Figure 22: Köhler clock detail. Top row, left to right: BAM, ground truth, and Carbajal et al. with joint training. Bottom row: Carbajal et al. without joint training, RAM, and the observation. Joint training refers to the kernel-estimation/deconvolution pipeline of Carbajal et al. (2023) .
MethodPSNR ↑\uparrowLPIPS ↓\downarrowCMMD ↓\downarrowFID ↓\downarrow
BAM16.140.7462.11312.88
BAM ⋆\star30.950.2360.1243.10
RAM8.960.8262.41377.62
RAM ⋆\star30.440.4240.63103.79
SILO23.540.4460.83110.87
UD2M30.950.2830.4255.71
Appendix
Table 9: FFHQ JPEG restoration. Results over 64 images at JPEG quality factor Q=10Q=10 . FT denotes finetuning.
Figure 23: FFHQ JPEG restoration. Quality factor 10, with Gaussian noise of standard deviation 0.010.01 in [0,1][0,1] added before compression. All columns refer to the same clean image. Matched yellow boxes show regions magnified 4×4\times .
ProblemBAM PSNR(mean)BAM mean PSNRRAM ⋆\star
Deblurring32.4930.6132.31
Super-resolution31.5729.9732.36
Inpainting34.2032.9033.36
Demosaicing38.1036.4437.36
Compressed sensing36.1834.5833.74
Appendix
Table 10: Single-problem FFHQ results. PSNR (dB; ↑\uparrow ) over 64 images and 64 draws per image at σy=0.025\sigma_{y}=0.025 . PSNR(mean) evaluates the mean of the 64 draws; mean PSNR averages the PSNR of those draws. Bold values mark the best in each row.
Figure 24: Posterior draws for a single problem. One FFHQ image per problem at σy=0.025\sigma_{y}=0.025 . The three displayed draws illustrate diversity rather than typical random draws. The mean uses all 64 draws. Matched yellow boxes show regions magnified 4×4\times .
ModelLatency (ms)Throughput (qps)PSNR ↑\uparrowLPIPS ↓\downarrow
Baseline280.563.5622.490.400
INT8-Optimized79.9112.5121.200.427
Appendix
Table 11: Effect of INT8 quantization on inference efficiency and reconstruction quality. We report latency, throughput, PSNR, and LPIPS for the baseline and INT8-optimized models. Bold values mark the best in each column.
Figure 25: Qualitative comparison of the baseline and INT8-optimized models on Gaussian deblurring. Reconstruction results are shown for the original baseline (left) and the INT8-optimized model (right) on test images from the LSDIR dataset. The close visual agreement between both outputs indicates that the proposed INT8 optimization substantially improves inference efficiency while preserving the reconstruction quality of the baseline model.

Explore similar work

CardsList

  1. Does AI Understand Imaging? A Systematic Benchmark of Agentic AI for Computational Imaging Tasks

    Jul 8, 2026Ethan Chung, Chuanjun Zheng, Jasper Tan +3Multi-Image Editing BenchmarksMultimodal Benchmarks

    University of Hawaii at Manoa · Glass Imaging

  2. Learning Normalized Energy Models for Linear Inverse Problems

    May 15, 2026Nicolas Zilberstein, Santiago Segarra, Eero Simoncelli +1Bayesian Inverse ProblemsDenoising Diffusion Probabilistic Model

    Rice University, Houston, TX, USA · Flatiron Institute, New York, NY, USA · New York University, New York, NY, USA

  3. Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer

    Nov 27, 2025Z-Image Team, Huanqia Cai, Sihan Cao +21Multi-Reference Image GenerationImage Generation

    Z-Image Team, Alibaba Group