acceptodds
Withdrawn from ICLR 2027

Coordinating Bits and Priors: Compact and Faithful Decoding for Generative Image Compression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative image compression methods have recently shown strong perceptual quality at ultra-low bitrates. Pretrained generative models provide strong natural-image priors for ultra-low-rate image compression, but they are designed for image synthesis and generation rather than reconstructing a particular source from a decoded bitstream. This motivates two considerations: different stages of the generative prior contribute unequally to reconstruction, and the decoded evidence needs to be preserved as synthesis progresses toward pixels. We propose CFIC, a compact and faithful generative image compression framework built around these observations. A training-free analysis reveals that reconstruction importance varies substantially across stages, motivating asymmetric compaction of the one-step diffusion prior rather than uniform shrinking. Multi-scale evidence recovered from the same quantized representation is then coordinated with corresponding synthesis stages, allowing source-specific information to complement the generative prior without additional side information. Finally, we alleviate the decoding bottleneck inherited from the pretrained model by preserving this evidence hierarchically along the path to pixels. Experimental results demonstrate that our proposed method achieves state-of-the-art perceptual quality with a compact model size and lower computational cost.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Learned Image Compression, Generative Image Compression, Diffusion Model, Computer Vision

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aSGdOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CFIC, a generative image compression framework that aims to reduce model complexity while improving source-specific reconstruction quality at ultra-low bitrates. The method consists of three main components: asymmetric pruning of a pretrained one-step diffusion U-Net based on reconstruction sensitivity, Bit-Prior Coordination Modules (BPCM) that inject decoded multi-scale features into the generation process, and Latent Interface Steering (LIS) that preserves source information throughout the final image decoding stages.

Critical strengths

The main strength of this paper is its strong quantitative perceptual compression performance with a substantially reduced model size. The results on Kodak, DIV2K, and CLIC2020 show improvements in LPIPS and DISTS compared with several diffusion-based compression methods. The reduction from a large pretrained diffusion model to a 267M-parameter system is also meaningful.

Another positive aspect is the asymmetric pruning analysis. Table 3 provides a relatively convincing comparison against uniform pruning under similar parameter budgets and recovery training conditions. The asymmetric model retains reconstruction quality close to the unpruned model, suggesting that reconstruction sensitivity can be useful for allocating model capacity more effectively.

Critical weaknesses

My first concern is the limited methodological novelty and the insufficient validation of the claimed faithful reconstruction mechanism. The proposed BPCM and LIS modules mainly rely on multi-scale feature injection, cross-attention, residual connections, and trainable decoder adapters, which are closely related to existing conditioning and feature fusion techniques. Although Table 2 shows that removing these modules degrades performance, it does not establish whether the specific designs are necessary compared with simpler alternatives, such as direct feature addition, concatenation, or standard cross-attention with comparable parameters. Similarly, the so-called innovation residual in BPCM is constructed from two learned feature projections, but the paper does not clearly demonstrate that it actually captures source-specific information missing from the generative prior.

More importantly, I am not convinced by the qualitative results. In Appendix D (Figs. D.1–D.4), I find that CFIC's reconstructions look visually worse than those of AEIC-SE, particularly in local structures and fine details. The results do not show a convincing visual advantage for CFIC, despite the better aggregate perceptual metrics reported in the paper. Since "faithful decoding" and preservation of source-specific information are central claims, I consider this discrepancy important. I would expect stronger evidence that the proposed method actually preserves source content more faithfully, rather than only improving selected perceptual metrics.

My second concern is the practical efficiency and the limited comparison with recent lightweight generative compression methods. Although CFIC significantly reduces parameters and MACs compared with large diffusion-based codecs, the actual latency improvement over AEIC-SE is almost negligible: 285.9 ms versus 289.7 ms in Table 1, despite the much larger reduction in parameter count. This makes the practical efficiency advantage less convincing than the paper suggests.

Moreover, recent works such as CoD-Lite (Jia et al., ICML 2026) have specifically focused on lightweight and real-time generative image compression, reporting a much smaller model and fast encoding and decoding. Such methods are highly relevant to CFIC's compactness and efficiency claims, but are missing from the evaluation and discussion. I am not suggesting that their reported runtimes can be directly compared across different hardware and evaluation protocols. However, without considering these recent lightweight alternatives, it is difficult to judge how competitive CFIC really is in terms of the overall quality–complexity–latency trade-off.

In addition, the pruning sensitivity analysis is conducted on the small Set14 dataset, while the main pruning comparison in Table 3 is performed at 0.0549 bpp, outside the main ultra-low bitrate range emphasized in Fig. 4. It remains unclear whether the same capacity allocation is consistently optimal across different image contents and bitrate settings.

Questions for authors
  1. Novelty of BPCM and LIS: What are the main technical advantages of BPCM and LIS over standard multi-scale conditioning or decoder feature injection? Have the authors considered simpler alternatives, such as direct feature fusion or standard cross-attention, under comparable parameter budgets? In particular, is there evidence that the proposed innovation residual captures information that is not already available in the generative features?

  2. Qualitative reconstruction quality: In Appendix D (Figs. D.1–D.4), CFIC's reconstructions appear visually less convincing than AEIC-SE to me, despite the reported improvements in perceptual metrics. Could the authors explain this discrepancy? Are matched-bitrate visual comparisons or per-image fidelity measurements available to better support the claim of faithful decoding?

  3. Practical efficiency and missing comparisons: CFIC reports substantially fewer parameters than AEIC-SE, but almost identical end-to-end latency (285.9 ms vs. 289.7 ms). What limits the actual speed improvement? Also, why are recent lightweight generative compression methods such as CoD-Lite (Jia et al., ICML 2026) not included or discussed? A comparison of the quality–complexity–latency trade-off would help clarify CFIC's practical advantages.

  4. Generalization of the pruning strategy: The sensitivity analysis is performed on Set14, and Table 3 evaluates the pruning strategies at a single bitrate. Have the authors examined whether the stage-wise sensitivity ranking and the advantage of asymmetric pruning remain consistent across the main ultra-low bitrate range and different image contents?

4 comments

Thank you for examining the pruning comparison and the details in the reconstruction crops. We added a trainable entrance control and measured total coding time under shared settings.

BPCM and LIS. BPCM routes decoded evidence during U-Net synthesis, and LIS keeps it accessible in later image-decoder layers. Both draw from the same coded representation. BPCM Direct keeps the projected difference E(e) − D(h) and bypasses attention. Its penalties support attention routing with this value construction, which is affine-equivalent to projecting the concatenated evidence and context. The trainable four-channel entrance still incurs a 25.40% DISTS BD-rate penalty, and removing hierarchical injection incurs a 21.92% penalty. These comparisons support the wide entrance and continued conditioning after the first projection.

Kodak BD-rate (%) relative to full CFIC

VariantPSNRLPIPSDISTS
Without BPCM+5.56+5.74+8.50
BPCM Direct+2.87+2.63+2.97
Without hierarchical LIS injection+8.23+12.94+21.92
Frozen 4-ch entrance+2.83+15.27+25.77
Frozen repeated entrance+14.06+13.42+25.58
Trainable 4-ch entrance (new)+4.16+14.79+25.40

Local reconstruction quality. A lower image-level DISTS can coexist with less convincing detail in a crop. Figs. D.1–D.4 show the same crop regions at nearby, unequal rates. Fig. 5 reports image-level bpp and DISTS, and CFIC has lower values on both measures in each of the three examples shown. Fig. 4 reports the dataset-level RD results, and Fig. C.2 locates intermediate feature updates around structured regions.

Fig. 5 exampleAEIC-SE bpp / DISTSCFIC bpp / DISTS
Rock texture0.0120 / 0.07690.0105 / 0.0664
Foliage0.0151 / 0.14900.0133 / 0.1014
Indoor signage0.0191 / 0.06490.0184 / 0.0510
Appendix exampleDatasetAEIC-SE bppCFIC bpp
D.1 PortraitCLIC20200.01580.0154
D.2 FixturesCLIC20200.00970.0092
D.3 BuildingsDIV2K0.00930.0090
D.4 GlasswareDIV2K0.01480.0136

Practical efficiency. At 1024 × 1024 on the RTX 4090, CFIC uses fewer parameters and MACs than AEIC-SE, with similar end-to-end latency. Runtime also depends on operator execution and coding overhead.

Method (Table 1)ParamsTMACsEnd-to-end (ms)
DiT-IC1.05B7.61524.8
AEIC-SE930M2.82289.7
CFIC267M2.40285.9

The new encoding-plus-decoding measurements use full-graph compilation, FP32 with TF32 enabled, and SDPA for both codecs. The U-Net uses tile size 128 with overlap 32. The VAE uses tile size 128 with halo 11. The second row tests a representative DIV2K/CLIC2020 resolution.

InputCFIC (ms)AEIC-SE (ms)Time reduction
Kodak79.1484.946.83%
1408 × 20481203.761252.193.87%

CoD-Lite. CoD-Lite has about 83M parameters and reports fast coding. We compare RD using five CFIC points and four CoD-Lite points per reported entry. CFIC's largest rate savings are on DISTS.

CFIC relative to CoD-Lite, cubic BD-rate (%)

DatasetPSNRMS-SSIMLPIPSDISTSFID
Kodak−23.92−7.04−16.68−48.65—
DIV2K−8.90—−31.37−71.72−65.78
CLIC2020−14.71—−39.43−74.01−65.21

Capacity allocation across bitrates. We use the Set14 sensitivity profile to guide topology selection, distill the selected topology once, and adapt the codec and reconstruction weights for each rate. Table 3 compares the two allocations at one Kodak operating point (0.0549 bpp). Asymmetric pruning preserves more reconstruction quality before recovery and retains its advantage afterward. Fig. 4 reports the RD performance of this shared topology after rate-specific adaptation on all three datasets.

Model (Table 3)PSNR (dB)LPIPSDISTS
Unpruned teacher (859M)24.710.1070.067
Asymmetric (267M), before recovery24.590.1160.085
Uniform (273M), before recovery16.690.3570.264
Asymmetric (267M), after recovery24.670.1080.069
Uniform (273M), after recovery23.930.1250.076

Thank you for the clarification and additional experiments. I appreciate the trainable four-channel entrance control, which strengthens the evidence for LIS, and the added RD comparison with CoD-Lite. However, my main concerns remain. The original latency is approximately 286 ms, and the additional measurements show only 6.83% and 3.87% time reductions under their respective settings. These gains are too modest to establish a meaningful practical speed advantage. CoD-Lite explicitly targets lightweight models and very fast decoding, which I consider a particularly promising direction for practical generative compression. Comparing RD performance without a corresponding latency comparison does not establish CFIC’s competitiveness in this trade-off. RDVQ is also an important missing baseline for evaluating reconstruction quality at a comparable model size. Furthermore, the image-level metrics do not resolve my concerns about local structural fidelity, and the response does not establish whether asymmetric pruning consistently outperforms uniform pruning across the main ultra-low-bitrate range. I therefore maintain my rating.

Thank you for your suggestion!

Our method is slower than CoD-Lite in encoding/decoding speed, but it significantly outperforms CoD-Lite on BD-Rate; we consider this a trade-off between speed and bitrate efficiency. For the experiments of CoD-Lite (additional), RDVQ and asymmetric pruning at ultra-low bitrates, we are conducting further experiments and will report the results.

Thank you for taking the time to respond to our comment, share your remaining concerns, and point out the experiments missing from our first response. Your feedback has helped us strengthen the study, and we have worked to address each concern below.

Latency and compactness. The end-to-end times below include encoding and decoding for 1024 × 1024 inputs on an RTX 4090.

MethodParametersLatency (ms)TMACs
CoD-Lite80M71.70.58
RDVQ252M1134.94.19
CFIC267M285.92.40

CoD-Lite offers the shortest coding time at 71.7 ms. CFIC takes 285.9 ms and uses fewer bits at matched PSNR, LPIPS, and DISTS across all three datasets in our earlier comparison. The DISTS BD-rate savings range from 48.65% to 74.01%. In this comparison, CFIC trades higher latency for better bitrate efficiency at a given reconstruction quality.

By “compact,” we mean that CFIC reduces the size of the pretrained one-step diffusion prior. We use reconstruction sensitivity to retain capacity in stages that matter most for reconstruction and prune less sensitive stages more aggressively. This reduces the full codec from 859M to 267M parameters, with reconstruction quality remaining close to the unpruned model at the operating point in Table 3.

RDVQ comparison. At a similar parameter count, CFIC has 74.8% lower end-to-end latency than RDVQ and the following BD-rates relative to RDVQ.

DatasetPSNRMS-SSIMLPIPSDISTSPatch-FID
Kodak−48.26%−15.38%−28.10%−5.49%—
DIV2K−37.00%—−39.94%−39.96%−54.00%
CLIC2020−33.30%—−41.43%−44.33%−71.17%

Pruning at ultra-low bitrates. Across five operating points per variant (0.006–0.044 bpp), asymmetric pruning yields cubic BD-rates of −10.26% (PSNR), −17.61% (LPIPS), and −23.54% (DISTS) relative to uniform pruning. Below are original BD points:

Pruning variantBPPPSNR (dB) ↑LPIPS ↓DISTS ↓
Asymmetric Pruned (CFIC, Ours)0.0066820.080.29080.1382
Asymmetric Pruned (CFIC, Ours)0.0131121.320.21730.1115
Asymmetric Pruned (CFIC, Ours)0.0237822.550.16160.0883
Asymmetric Pruned (CFIC, Ours)0.0322023.170.13700.0804
Asymmetric Pruned (CFIC, Ours)0.0431023.770.11910.0703
Uniform Pruned0.0061419.930.32240.1504
Uniform Pruned0.0129221.140.23810.1218
Uniform Pruned0.0239122.250.17880.0984
Uniform Pruned0.0324222.870.15090.0893
Uniform Pruned0.0435123.430.13530.0816

Local structural fidelity. We will use “source-conditioned reconstruction” in place of “faithful decoding” and tie our quality claims to the reported metrics.

Thank you!

Reviewer_yHAhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CFIC, an ultra-low-bitrate generative image compression framework combining asymmetric compaction of a one-step diffusion prior, multi-scale bit-prior coordination, and adaptation of the final image-decoder interface. A stage-wise reconstruction-sensitivity diagnostic motivates pruning deeper, lower-resolution stages more aggressively. Bit-prior coordination modules reuse features derived from the coded latents at multiple synthesis stages, while Latent Interface Steering combines a trainable wide decoder entrance with hierarchical injection. Experiments on Kodak, DIV2K, and CLIC2020 report strong perceptual rate-distortion performance with a 267M-parameter model. Recovery and component ablations support the utility of nonuniform capacity allocation and continued source conditioning.

Critical strengths
  1. The method has a coherent rationale for compact generative decoding: retain capacity in reconstruction-sensitive stages and keep decoded source features accessible throughout synthesis. Reusing features from the same coded latents in both the U-Net and image decoder provides continued source conditioning without an additional coded feature stream.
  2. The parameter–quality trade-off is useful. CFIC uses 267M parameters versus AEIC-SE's 930M; Table 1 reports DISTS BD-rate savings of 44.81%/42.05% relative to DiT-IC on DIV2K/CLIC2020. Table 3 provides a controlled pruning comparison: at 0.0549 bpp on Kodak, asymmetric pruning reaches 24.67 dB/0.108 LPIPS versus uniform pruning's 23.93 dB/0.125, close to the unpruned teacher's 24.71 dB/0.107.
Critical weaknesses
  1. The benefits of the specific method choices are not sufficiently isolated. Although Table 2 includes frozen four-channel and repeated-wide entrances, it lacks a trainable four-channel control to separate widening from unfreezing. In BPCM, the independently learned 1×1 projections (Appendix A.3) make E(e)−D(h) equivalent to a projection of [e,h]; removing the whole branch does not establish the benefit of this particular value construction over evidence-only values. For pruning, Table 3 supports nonuniform allocation, but the rule mapping sensitivity scores to the final configuration is underspecified (§3.2, Appendix C.1). The distinction from PocketSR's depth-dependent pruning and annealing also needs clearer explanation.
  2. The comparison needs clearer coverage and qualification. Figure 4 omits RDEIC on DIV2K despite reporting its BD-rates in Table 1, and shows ARPC only on Kodak without explaining the dataset restrictions. These omissions need explanation to establish which methods and datasets the SOTA claim covers. The fixed 1024×1024 timing comparison is useful, but CFIC and AEIC-SE have nearly identical latency (285.9/289.7 ms). The paper acknowledges this comparable latency, but does not establish whether the efficiency comparison holds for original-size test images under tiled inference.
Questions for authors
  1. Can you unfreeze the entrance in the existing four-channel baseline and add a BPCM control using evidence-only values E(e), keeping the other settings unchanged? Please also specify how the sensitivity profile determines the final pruning configuration and how this allocation differs from PocketSR's depth-dependent pruning strategy.
  2. Please report CFIC and AEIC-SE's average encoding and decoding times on original-size Kodak, DIV2K-validation, and CLIC2020 test images, with the Figure 4 inference pipeline. One representative low-rate checkpoint per method is sufficient; specify tile size, overlap, and timing variability.
  3. Please provide the RDEIC DIV2K operating points used in Table 1 and explain their absence from Figure 4. Is ARPC's omission on DIV2K/CLIC2020 due to resolution restrictions? Clarify the evaluation exceptions and qualify the SOTA claim accordingly.
1 comment

Thank you for suggesting a trainable four-channel control and asking how we measured inference time. We ran both comparisons.

Entrance width and trainability. On Kodak, unfreezing the four-channel entrance changes the perceptual BD-rate penalties only slightly. The trainable four-channel model still needs 14.79% more bits at matched LPIPS and 25.40% more at matched DISTS than full CFIC. This comparison supports the learned wide entrance. The frozen repeated entrance collapses inputs with the same channel-group sum. Learning independent wide weights removes that constraint.

Kodak BD-rate (%) relative to full CFIC

VariantPSNRLPIPSDISTS
Frozen 4-ch entrance+2.83+15.27+25.77
Frozen repeated entrance+14.06+13.42+25.58
Without hierarchical LIS injection+8.23+12.94+21.92
Trainable 4-ch entrance (new)+4.16+14.79+25.40

BPCM value and routing. We agree that E(e) − D(h), with independently learned projections, is affine-equivalent to projecting the concatenated evidence and context. BPCM routes this context-relative update with synthesis queries and native-skip keys. BPCM Direct keeps the projected difference and bypasses attention. Its BD-rate penalties support attention routing with this value construction.

Kodak BD-rate (%) relative to full CFIC

VariantPSNRLPIPSDISTS
Without BPCM+5.56+5.74+8.50
BPCM Direct+2.87+2.63+2.97

Sensitivity and PocketSR. In Appendix C.1, we mask half the channels in one stage at a time, keep the decoded latents and other parameters fixed, and measure changes in LPIPS and DISTS on Set14 before recovery. The profile guides the allocation within the model budget and channel-interface constraints. Fig. C.1 shows the selected configuration. We protect Down0, Down1, Up2, and Up3, remove about 94% of Down2 and 82% of Up1 parameters, and remove Down3, the middle block, and Up0 entirely. Table 3 compares this allocation with uniform pruning before and after matched recovery. We adopt PocketSR's heavy-to-light annealing and feature distillation. CFIC removes deep blocks and reconnects the retained path with residual links. BPCM and LIS condition reconstruction on entropy-decoded features under a bitrate objective.

End-to-end timing. Both codecs use full-graph compilation, FP32 with TF32 enabled, and SDPA. The U-Net uses tile size 128 with overlap 32. The VAE uses tile size 128 with halo 11. Both rows report encoding plus decoding. The second row tests a representative DIV2K/CLIC2020 resolution.

InputCFIC (ms)AEIC-SE (ms)Time reduction
Kodak79.1484.946.83%
1408 × 20481203.761252.193.87%

Table 1 measured 285.9 ms for CFIC and 289.7 ms for AEIC-SE at 1024 × 1024 on an RTX 4090, including entropy encoding and decoding. The new timing configuration enables VAE tiling. The submitted reconstruction protocol leaves the image decoder untiled by default (Appendix B.2).

Benchmark coverage. RDEIC's DIV2K measurements contribute to Table 1. Fig. 4 omits the DIV2K curve. The table gives aggregate BD-rates relative to DiT-IC over the shared measured quality range.

DatasetRDEIC LPIPS (%)RDEIC DISTS (%)CFIC LPIPS (%)CFIC DISTS (%)
Kodak+52.82+121.37−7.50−9.26
DIV2K+51.76+127.28−19.39−44.81
CLIC2020+65.97+260.04−18.93−42.05

ARPC reports 1024 × 1024 center crops on DIV2K and CLIC2020. Our protocol uses full images, and we evaluate ARPC on Kodak. The perceptual RD claim applies to the method–dataset pairs evaluated under the stated protocols.

Reviewer_XyAfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CFIC for generative image compression. The authors prune the diffusion U-Net based on how much each stage affects image reconstruction. This reduces the model size and decoding latency. They also inject multi-scale features recovered from the bitstream during generation. Latent Interface Steering (LIS) keeps these features available during image decoding.

Critical strengths
  • The method considers both the size of the generative prior and how to use information from the bitstream. It recovers multi-scale features from the same bitstream and uses them at different decoding stages. The overall design is clear.
  • The authors use reconstruction sensitivity to guide asymmetric U-Net pruning. They keep more capacity in stages that matter more for reconstruction.
Critical weaknesses
  • The pruning setup across bitrates is unclear. Is the pruning structure selected at one fixed bitrate and then fine-tuned for other bitrates? Or is a different structure selected for each bitrate? The bitstream carries different amounts of information at different bitrates, so the importance of each U-Net stage may change. The author should give more detailed explanation about this.
  • The paper reports decoding latency, but didn't provide the time for sensitivity analysis and pruning. How about GPU-hours for the full pruning process?
  • Table 2 compares LIS with several frozen decoder ablations. It does not compare LIS with fine-tuning the full decoder or fine-tuning decoder with LoRA. Why choose LIS but not finetune the decoder? The authors should compare LIS with the finetuned decoder and report their trainable parameters and training costs.
Questions for authors

see weakness

1 comment

Thank you for recognizing the asymmetric pruning design and the reuse of bitstream features throughout reconstruction. Your questions help clarify how we construct and adapt the compact model.

Pruning across bitrates. We select the compact topology once from the fully adapted teacher, recover the student, and adapt its codec, low-rank adapters, and coordination modules for each target rate. The block layout and channel widths stay fixed. This shares the heavy-to-light transfer across operating points while allowing the reconstruction model to adapt to the amount of transmitted information. Figure 4 reports the resulting rate-specific models across the evaluated bitrate range.

Construction and rate adaptation. Appendix B.1 gives the optimization schedule below, with an effective batch size of 8 throughout. The 20K-step heavy-to-light transfer is shared across rates. Each target rate then uses 20K steps of adaptation with 512 × 512 crops and 5K steps of high-resolution fine-tuning with 1024 × 1024 crops. Training-only heavy branches are removed and low-rank adapters are merged for deployment.

Training stageOptimization stepsDetails
Full-model teacher pretraining250KMLIC-Train-100K; 512 × 512 crops
Asymmetric compaction and transfer20K5K upsampling-path annealing + 5K downsampling-path annealing + 10K recovery
Rate-specific fine-tuning20K per rate512 × 512 crops
High-resolution fine-tuning5K per rateDF2K and CLIC Train; 1024 × 1024 crops

Basis for the capacity allocation. Appendix C.1 measures the reconstruction change caused by masking channels in one stage at a time, with decoded latents and the remaining stages fixed. The observed sensitivity motivates retaining more capacity in shallow and high-resolution stages. Table 3 tests the selected allocation against uniform pruning with matched lightweight operators and recovery, as reproduced below.

Controlled pruning comparison on Kodak at 0.0549 bpp (Table 3).

Model / stageParametersPSNR (dB)LPIPSDISTS
Unpruned teacher859M24.710.1070.067
Asymmetric student, before recovery267M24.590.1160.085
Uniform student, before recovery273M16.690.3570.264
Asymmetric student, after recovery267M24.670.1080.069
Uniform student, after recovery273M23.930.1250.076

Before recovery, the asymmetric topology already preserves much more reconstruction quality than uniform pruning. After matched recovery, it exceeds the uniform student by 0.74 dB PSNR and reduces LPIPS and DISTS by 0.017 and 0.007. Relative to the teacher, the remaining differences are 0.04 dB PSNR, 0.001 LPIPS, and 0.002 DISTS.

LIS versus full-decoder fine-tuning. We tested full-decoder fine-tuning while retaining the wide entrance and removing hierarchical LIS injections. Relative to full CFIC, this variant improves PSNR BD-rate by 2.37%, but increases LPIPS and DISTS BD-rates by 10.13% and 18.17%. This result supports LIS for the perceptual RD objective: its controllers give later decoder layers access to source features as their reconstruction states evolve. The pretrained decoder body remains frozen, while the wide entrance and residual controllers are trained.

Decoder adaptation on Kodak24. Values are cubic BD-rate changes relative to full CFIC; positive values indicate a rate penalty.

VariantPSNRLPIPSDISTS
Frozen four-channel conv-in (Table 2)+2.83%+15.27%+25.77%
Frozen repeated conv-in (Table 2)+14.06%+13.42%+25.58%
Hierarchical LIS injection removed (Table 2)+8.23%+12.94%+21.92%
Full-decoder fine-tuning without hierarchical LIS injections−2.37%+10.13%+18.17%
Trainable four-channel conv-in+4.16%+14.79%+25.40%

The trainable four-channel control also performs worse than full CFIC across all three metrics. Together, these comparisons support learning a wide entrance and maintaining source conditioning at later decoder depths.

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

Related papers

Discussion (0)

Sign in to comment.