Neural image codecs face a fundamental trade-off between pixel fidelity and perceptual quality. Distortion-optimized codecs preserve pixel-wise fidelity but produce blurred reconstructions at low bitrates, whereas generative codecs leverage learned priors to improve perceptual quality but tend to exhibit saturating PSNR at higher bitrates and require separate models for different bitrate targets. We propose PerQ, which reframes perceptual reconstruction as inverse generative modeling with bounded support. Specifically, we define the quantization cell as the set of latents that yield the same transmitted symbols and derive a transport bound showing that reconstructions from this cell remain within a rate-determined radius of the codec output. This motivates performing generative reconstruction within the codec-induced quantization cell, rather than sampling from an unconstrained image prior. PerQ implements this principle by attaching a rate-conditioned flow-matching compensator to a frozen, multi-rate distortion-optimized backbone, enabling perceptual enhancement while preserving the backbone's coding performance. Experiments on Kodak and CLIC2020 show that PerQ achieves perceptual performance comparable to generative codecs such as MS-ILLM in LPIPS and FID, while maintaining competitive PSNR relative to distortion-optimized codecs such as ELIC. A single model checkpoint supports a wide bitrate range with fast decoding, and the compensator provides consistent gains across three different backbones. The implementation is available at https://jasminepp.github.io/PerQ/.
Figures & tables
Figure 1: Balance between Perceptual quality and pixel fidelity. (Left) Visual comparison at comparable bitrates, with PSNR (dB) / LPIPS. (Right) BD-rate of PSNR against BD-rate of LPIPS on CLIC2020-test, both relative to VTM intra, lower is better).
Kodak : BD-rate vs. VVC (%) ↓
CLIC2020-test : BD-rate vs. VVC (%) ↓
Image codecs
Single-model multi-rate
Dec. time (ms/Mp) ↓
Dec. kMACs/px ↓
PSNR
LPIPS
DISTS
FID
PSNR
LPIPS
DISTS
FID
Conventional codec
VTM intra Bross et al. (2021)
–
–
–
0
0
0
0
0
0
0
0
VAE-based codecs
ELIC He et al. (2022a)
206
332
−2.0
−2.1
+2.4
+7.3
−1.9
−1.3
+9.4
−13.8
EVC Wang et al. (2023)
✓
57
291
+2.0
−9.0
−5.4
−4.0
+3.7
−8.7
−2.3
−6.2
Table 1: Positioning of image codecs across compression paradigms. Each row reports single-model multi-rate operation, decoding complexity and BD-rate against VTM intra on the Kodak Eastman Kodak Company () and CLIC2020-test CLIC (2020) datasets.
Figure 2: PerQ overview. Top: the compression pipeline. Bottom-left: architecture of the flow-matching based compensator. Bottom-right: illustrative 2D slice of latent space. The cell is centered at the quantized latent y^ and contains the source latent y by the quantization bound ( 3 ). The compensator generates y~ within the cell, targeting the high-perceptual-quality field ( blue ) while remaining inside the low-distortion field ( orange ).
Figure 3: Symbol-space distribution of the backbone quantization error ysym−y^sym on Kodak at four rate points (log density). The sub-cell ±m/2 (red, m=0.1 ) contains most of the error at low rates, and the prior N(0,σ2) with σ=m/12 (gold) is set to the scale of this sub-cell.
Configuration
PSNR ↑
MS-SSIM ↑
LPIPS ↓
DISTS ↓
FID ↓
Loss setting
v1.1 LPIPS instead of E-LPIPS
+1.92%
+1.16%
+2.85%
+1.39%
+5.13%
v1.2 E-LPIPS weight 0.5
+0.24%
+0.97%
+1.38%
-2.57%
-1.60%
v1.3 E-LPIPS weight 0.25
+0.61%
+2.02%
+3.49%
+1.15%
+7.20%
Flow matching steps
v2.1 1 step
-1.06%
-1.83%
+7.96%
+5.63%
+14.75%
Table 2: PerQ design ablation on CLIC2020-test. BD-rate against the default setting (DCVC-RT intra backbone, 31.0M parameters (the compensator), 4 steps, λFM=104 and trained with the patch-GAN and E-LPIPS perceptual loss (weight 1.0 )).
Bound scale m
PSNR ↑
MS-SSIM ↑
LPIPS ↓
DISTS ↓
FID ↓
PSNR min
PSNR max
0.05
+13.77%
+16.20%
-60.56%
-76.61%
-91.87%
24.88
39.16
0.10 (default)
+18.24%
+20.60%
-67.84%
-79.68%
-91.51%
24.76
38.81
0.20
+23.23%
+24.40%
-70.79%
-82.53%
-92.39%
24.41
38.47
unbounded
+28.58%
+25.38%
-64.45%
-81.73%
-92.22%
24.47
38.11
Table 3: Bound scale ablation on CLIC2020-test. BD-rate against the unmodified backbone for the bound scale m in ( 10 ). PSNR min /PSNR max report the average PSNR at the lowest/highest rate point.
Figure 4: Rate–distortion curves on Kodak (top) and CLIC2020-test (bottom). Panels report PSNR (RGB), LPIPS (AlexNet), DISTS, and FID across bitrates. PerQ ( red star ) traces a single curve from one trained checkpoint by varying the rate parameter at the decoder. Solid markers denote multi-rate or distortion-optimal baselines; dashed markers denote per-rate trained generative codecs.
Kodak
CLIC2020-test
bpp
PSNR
LPIPS
DISTS
FID
bpp
PSNR
LPIPS
DISTS
FID
DCVC-RT backbone Jia et al. (2025) (default)
Backbone only
0.010–0.87
−10.6
−12.7
−5.7
−10.4
0.009–0.54
−11.1
−7.3
+3.5
−42.2
+ PerQ compensator
0.010–0.87
−1.6
−57.4
−63.2
−67.9
0.009–0.54
+2.7
−70.6
−77.4
−91.9
ELIC backbone He et al. (2022a)
Backbone only
0.061–0.91
−4.9
−13.8
−5.2
−5.5
0.045–0.56
−5.5
−8.5
+4.9
−28.4
Table 4: PerQ compensator on three frozen backbones. The same compensator, bound and training recipe are attached to three variable-rate backbones (retrained), each covering 64 rate points with one checkpoint. BD-rate (%) against VVC intra.
Figure 5: Qualitative comparison. Source crop (left) compared with reconstructions from a distortion-optimized VAE-based codec ( grey ), PerQ ( red , ours), and two generative codecs, GAN-based and diffusion-based ( blue ), at similar bitrates with the error maps.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Codec
Single-model multi-rate
bpp range
PSNR
LPIPS
DISTS
FID
KID
Conventional codec
VTM intra [ 10 ]
0.113–2.316
0
0
0
0
0
VAE-based codecs
ELIC [ 20 ]
0.044–0.865
−2.0
−2.1
+2.4
+7.3
+3.8
EVC [ 56 ]
✓
0.329–0.983
+2.0
−9.0
−5.4
−4.0
−7.5
DCVC-RT-Intra [ 25 ]
✓
0.031–0.913
−11.2
−8.7
−2.3
−1.8
−2.8
Appendix
Table 5: BD-rate against VVC on Kodak (%, lower is better) with bitrate ranges and KID. Same anchor, fitting method and N/A rule as Table 1 . The bpp range covers the rate points evaluated for each codec.
Codec
Single-model multi-rate
bpp range
PSNR
LPIPS
DISTS
FID
KID
Conventional codec
VTM intra [ 10 ]
0.076–1.717
0
0
0
0
0
VAE-based codecs
ELIC [ 20 ]
0.032–0.538
−1.9
−1.3
+9.4
−13.8
−31.6
EVC [ 56 ]
✓
0.211–0.626
+3.7
−8.7
−2.3
−6.2
−11.9
DCVC-RT-Intra [ 25 ]
✓
0.024–0.562
−11.8
−2.5
+8.8
−30.1
−39.7
Appendix
Table 6: BD-rate against VVC on CLIC2020-test (%, lower is better) with bitrate ranges and KID. Same protocol as Table 5 .
Figure 6: Rate-PSNR and rate-KID curves on Kodak (top) and CLIC2020-test (bottom). KID ( ×103 , lower is better) is computed on the same 256×256 patches as FID.
Figure 7: Visual comparison between the backbone and PerQ on Kodak , with the DCVC-RT (top) and MLIC++ (bottom) backbones. Additional examples on CLIC2020-test are shown in Figure 8 (DCVC-RT) and Figure 9 (MLIC++).
Figure 8: Visual comparison between the backbone and PerQ on CLIC2020-test with the DCVC-RT backbone.
Figure 9: Visual comparison between the backbone and PerQ on CLIC2020-test with the MLIC++ backbone of the generality study ( Table 4 ).
Figure 10: Additional visual comparison between PerQ and baseline methods on CLIC2020-test [ 14 ] . Source crop (left) compared with reconstructions from our frozen backbone ( grey ), PerQ ( red , ours) at the same rate point, and two generative codecs, GAN-based and diffusion-based ( blue ), at similar bitrates. Below each reconstruction, the error map shows its absolute difference to the source on a shared color scale, and the numbers give the PSNR and LPIPS of the whole image.
Figure 11: Reconstructions across rates on CLIC2020-test [ 14 ] . Generative codecs ( blue ) and PerQ ( red ) at their own rate points, with the absolute error map below each crop. The PSNR of PerQ keeps rising with the rate, while Control-GIC and OSCAR stay below 25 dB.
Department of Information Engineering Shanghai University of Electric Power Shanghai, China · Department of Computer Science and Communication Engineering Waseda University Tokyo, Japan · Faculty of Engineering Institute of Science Tokyo Tokyo, Japan