acceptodds
Under review as a conference paper at ICLR 2027

PULSE: Unlocking Practical Image Compression on Single-Thread CPU

Abstract

Despite recent progress in learned image compression, existing methods remain computationally expensive on resource-constrained hardware, particularly CPUs. We introduce PULSE, a practical codec that enables (1) low-latency decoding on diverse hardware platforms with an ultra-low-complexity 5.2kMAC/pixel neural receiver, and (2) efficient bit-exact entropy coding} with an integer linear CDF predictor and a meta prior. To recover compression performance under this tight budget, we introduce an agentic evolution process guided by heuristic probes that iteratively improves the architecture through human-LLM collaboration. PULSE decodes a 1080p image in 126ms on a single CPU thread while achieving compression performance comparable to HM. After perceptual optimization, PULSE competes with larger perceptual codecs like MS-ILLM.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
neural image compression, lightweight neural networks, real-time compression

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

PULSE is an asymmetric variable-rate codec targeting inexpensive decoding. It combines a low-resolution decoder and renderer, integer linear CDF prediction with a meta prior, and architecture evolution guided by representation probes and human–LLM collaboration. A two-stage perceptual/ROI objective further targets textures, text, and faces. Validation includes multi-dataset rate–distortion results, 126-ms single-thread CPU decoding for 1080p at 5.2 kMAC/pixel, and cross-device entropy-coding checks and full-pipeline timings.

Critical strengths
  1. Full-pipeline single-thread timings include entropy coding, and the appendix documents backend and timing choices. Zero CDF-index and decoded-symbol mismatches across phone, laptop, and server configurations substantiate cross-platform entropy decoding. These measurements address practical issues that MAC counts alone cannot capture.

  2. Integer linear CDF prediction avoids a large deterministic neural probability-decoding path, while the meta prior recovers coding efficiency. Table 1 evaluates alternatives and their rate–distortion costs, making this component a specific, experimentally supported contribution to portable lightweight codecs.

Critical weaknesses

1.The human researcher interprets probes, changes the search space, selects candidates, and promotes models. The paper acknowledges that human and LLM contributions cannot be separated post hoc. Consequently, the 22-versus-1280 training-run comparison does not establish that the agentic process is more effective than human design using the same probes, or normalize total candidate-training and probe cost. The resulting architecture is useful, but the claimed search contribution remains difficult to attribute and reproduce as a method.

  1. Appendix Table 12 reports worse LPIPS-Alex BD-rate than MS-ILLM across all five datasets for both perceptual PULSE variants. Table 15 reports +17.8% FID BD-rate on CLIC test for perceptual PULSE, despite -4.1% on CLIC validation; PULSE-L does perform favorably on FID. Other metrics, including LPIPS-VGG and DISTS, also favor PULSE, so the evidence is mixed rather than uniformly negative. Without a blinded preference comparison, these conflicting indicators do not establish the broad subjective-quality interpretation, particularly for text and faces.
Questions for authors
  1. At equal total search cost, what advantage does human–LLM evolution provide over human design with the same probes? Does the 1280-run NAS comparison use comparable candidate-training budgets and search flexibility?

  2. How do you reconcile the LPIPS-Alex regressions and PULSE's CLIC-test FID result with the favorable VGG/DISTS metrics? Is there blinded human-preference evidence, especially for text and face regions?

2 comments

W1 & Q1: Equal-cost attribution of human, LLM, and NAS

We agree that a prospective human-only study using the same probes would be informative. However, a post-hoc human-only trajectory would be inherently biased because the researchers already know the successful directions and newly discovered search axes. A rigorous comparison would require independent experts who are unaware of the original trajectory, while also controlling for their expertise, time, and candidate-training budget. We therefore do not claim that human–LLM evolution outperforms expert-only design.

Instead, we perform an additional ablation study to determine the effects of human review and search flexibility. Each training run uses the same training budget and evaluation protocol. The full and human-free variants use the same probe budget and the same 22 full training runs.

WorkflowHumanNew search axes# Training RunsBD-rate
Full agentic evolution✓✓22-10.2%
LLM + probes✗✓22-8.7%
Agentic evolution✗✗22-7.1%
Hardware-aware NAS✗✗22-3.1%
Hardware-aware NAS✗✗1,280-7.5%
  • Human review further improves BD-rate by 1.5 percentage points, from -8.7% to -10.2%. Even without human review, the probe-guided LLM reaches -8.7%, outperforming exhaustive NAS at -7.5% with only 22 rather than 1,280 training runs.
  • Under the same predefined axes and the same 22-run budget, agentic evolution reaches -7.1%, compared with -3.1% for hardware-aware NAS. With only 22 runs, it also approaches the exhaustive NAS result of -7.5% obtained from 1,280 runs. Allowing the LLM to introduce new axes further improves the result to -8.7%; this ability to evolve the search space is a key advantage over fixed-space NAS.

W2 & Q2: Conflicting perceptual metrics and blinded preference

We agree that the evidence is metric-dependent: PULSE is worse than MS-ILLM on LPIPS-Alex and CLIC-test FID, but better on LPIPS-VGG and DISTS. These metrics capture different perceptual properties, so we will revise the paper to report this trade-off rather than claim uniform superiority.

To provide complementary subjective evidence, we conducted a blinded study comparing PULSE-Perceptual with MS-ILLM at comparable bitrates across all evaluated images. Twenty participants judged which reconstruction was visually closer to the ground truth or marked the two as indistinguishable.

RegionPULSE preferredMS-ILLM preferredIndistinguishable
Text37.5%6.0%56.5%
Face66.0%0.5%33.5%
Global23.9%11.1%65.0%

PULSE receives more preferences than MS-ILLM in all three categories, with the clearest advantage on text and face regions. For global images, both methods frequently produce reconstructions that are already very close to the ground truth at the tested bitrates, explaining the high indistinguishable rate. Overall, these results indicate comparable global perceptual quality, with a clearer preference for PULSE on ROI regions.

We sincerely appreciate your time and effort in reviewing our work and providing constructive feedback. As the rebuttal deadline approaches, we would be grateful if you could let us know whether our responses have adequately addressed your concerns or if any further clarification is needed. Thank you again for your time and consideration.

Reviewer_7g4UOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3(+1)
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

Learned image codecs exceed classical standards in rate–distortion but decode slowly on general-purpose CPUs, and their entropy decoding can fail across devices because it depends on floating-point arithmetic. The authors propose PULSE, an asymmetric codec that concentrates model capacity in the sender so that the receiver fits on a single CPU thread, and that confines the cross-platform determinism requirement to one integer linear projection that selects an entropy-coding CDF table, leaving the rest of the receiver on the device's neural backend. The claimed contributions are a receiver that decodes a 1080p image in about a tenth of a second on one thread at rate–distortion comparable to the HEVC reference decoder, bit-exact entropy coding across a phone, a laptop, and a server, a human–LLM design procedure presented as more sample-efficient than neural architecture search, and a perceptual variant with region-of-interest losses for faces and text.

Critical strengths
  1. The linear CDF-index decoder with the Meta Prior is a simple and specific solution to cross-platform determinism in learned entropy coding. Making the scale path independent of the mean path means only the CDF selection must be computed deterministically, and predicting the table index directly avoids the logarithm and exponential at run time. Table 1 isolates each step: relative to the proposed model in floating point, the integerized full entropy model takes 86 ms and is +5.7%+5.7\% in BD-rate at INT8 (+1.5%+1.5\% in floating point), whereas the proposed path takes 1.6 ms and is +0.04%+0.04\%, with zero CDF-index and symbol mismatches on three platforms (Table 6).

  2. Complexity is reported consistently. Sender and receiver parameters, MAC counts, and single-thread times are given for every neural baseline on the same CPU with the same rANS implementation (Appendix C.1), alongside timings of the classical codecs, with a per-component breakdown (Tables 4–5), five test sets, five metrics, and a comparison with PICO on its released reconstructions (Table 9).

  3. The low-resolution decoder with a 1/4-resolution renderer is supported by the ablation in Table 2, and the perceptual PULSE-L reaches lower FID, DISTS, and LPIPS-VGG BD-rates than MS-ILLM on all five test sets at 20 kMAC/pixel, compared with 675 kMAC/pixel for the MS-ILLM receiver (Tables 13–15).

Critical weaknesses
  1. The paper claims to occupy a "missing operating point" for practical image compression, but it does not state the deployment scenario, and the comparison set omits the codecs that would compete in either plausible scenario. If the scenario is delivery (encode once, decode many times on client CPUs), the relevant competitors are production decoders such as libjpeg-turbo, libwebp, libjxl for JPEG XL [e], dav1d for AVIF, and optimized VVC decoders [a]. HM-16.25 and VTM-17.0 are reference implementations, and their decoding times (98 ms and 189 ms) do not represent deployed decoders. VTM is −23.4%-23.4\% relative to HM on the same set, so an optimized VVC decoder at comparable single-thread speed would dominate PULSE in rate–distortion; the paper needs to measure this. In the delivery scenario, encoding cost is amortized, so overfitted codecs such as Cool-Chic and C3, which the introduction acknowledges reach receivers below 5 kMAC/pixel, belong in the left panel of Figure 1 with measured decoding times. The paper excludes them for their per-image encoding optimization and states that their sequential entropy models limit implementation efficiency (Section 7), but it does not measure that limitation. C3 matches VTM on CLIC2020 with fewer than 3 kMAC/pixel, and a reduced-complexity Cool-Chic decodes Kodak images (0.39 megapixels) on one CPU core in 100–268 ms at rate–distortion between HEVC and VVC [b]. Whether PULSE's speed at 1080p outweighs that rate–distortion gap is the comparison the paper's claim requires. If the scenario is capture on a device, the sender (45.35 kMAC/pixel) is the constraint. The paper reports 42 ms per 1080p encode on the Snapdragon 8 Elite NPU, but 920 ms on one CPU thread and 302 ms on the laptop CPU, so encoding at capture rates depends on an NPU; the relevant comparisons are then the hardware HEIC and JPEG encoders that phones capture with and asymmetric codecs with lightweight encoders [c, d]. In either case, the claim that PULSE-S is "faster than all compared codecs" depends on which codecs are compared. The only non-reference classical baseline, HEIC, is encoded in 4:2:0, while HM and VTM use 4:4:4, and the timing boundaries differ: HM and VTM timings use the native YUV interface and exclude RGB conversion, whereas the HEIC benchmark operates on in-memory RGB (Appendix C.1). The authors report that 4:4:4 was worse in their libheif/x265 setting (Appendix C.1). If PSNR is computed in RGB, as the RGB inputs suggest, this is unexpected at medium and high rates, and the 4:4:4 curve should be shown.

  2. Agentic evolution is presented as a primary contribution, but the evidence is a single search trajectory without controls. Table 2 compares 22 training runs at −10.2%-10.2\% with hardware-aware NAS (attributed to Tatwawadi et al., 2026, in Section 6.2) at 1280 runs and −7.5%-7.5\%, yet the paper does not state whether the NAS was re-run by the authors, in which search space, with which training schedule per candidate, or under which complexity constraint. The agentic process also introduced new design dimensions (renderer resolution at E2, grouped convolution at E5, FFN width at E7), and the paper does not state whether the NAS search space contained them, so the comparison conflates the design of the search space with the search procedure. Each candidate is trained once, and the late promotions (E5→E6, −0.5-0.5 points; E6→E7, −1.3-1.3 points) are of the same order as seed-to-seed variation in learned codecs, which the paper does not report. The paper also does not state on which images candidates were selected. If the selection set is CLIC Professional validation, it is also the headline test set; PULSE's BD-rate relative to HM is most favorable there (−2.9%-2.9\%) among the five test sets (Kodak +1.2%+1.2\%, CLIC test +0.3%+0.3\%). The authors state that the human and LLM contributions cannot be separated (Section 4.4), and the LLM is identified only by name (gpt-5.6-sol, Appendix D.1), so the procedure cannot be reproduced or attributed. Supporting the sample-efficiency claim would require repeated independent runs, a NAS baseline in the same search space and budget, and ablations without the LLM and without the probes. Otherwise, the architecture should be presented as the contribution and the procedure as an account of how it was found.

  3. The claim that PULSE (Perceptual) "competes with larger perceptual codecs like MS-ILLM" (Abstract) rests on metrics that disagree in sign across test sets and metrics. Relative to MS-ILLM, its FID BD-rate is −16.7%-16.7\% on Kodak, −4.1%-4.1\% on CLIC Professional validation, and −1.7%-1.7\% on DIV2K, but +12%+12\% on Tecnick, +17.8%+17.8\% on CLIC test, and +12.7%+12.7\% on the 300-image PICO subset (Tables 9 and 15), and its LPIPS-Alex BD-rate is worse by 25–32% on every set, while LPIPS-VGG, the larger term in its training loss, favors PULSE (Tables 12–13). With this disagreement, a human rating study, such as the one used to validate HiFiC, is needed to support the claim. The region-of-interest stage is motivated by text and face quality, but it is evaluated only with face and text PSNR (Table 8), and the paper does not state the evaluation images for Table 8 or the detector that produces the face and text regions. Section 5 states that the stage "jointly improves pixel fidelity, perceptual quality, and ROI quality, which is verified in Table 8", but in Table 8, removing the ROI losses improves LPIPS-VGG and FID BD-rate by 2.4% and 3.7%. Generative decoders can substitute characters or alter identity at low rates, so character error rate of a text recognizer and face verification accuracy on the decoded crops would test what the stage is designed to preserve.

  4. The "bit-exact entropy coding across platforms" result (Figure 1, Table 3 caption, Section 8) covers less of the receiver than the placement of the "Two-step entropy model" inside the Entropy Coding block of Figure 2 suggests. By Equations (3) and (5), the rANS decoder uses only the CDF-table index computed from z^\hat{z} by the integer projection, and Appendix A.3 states that entropy coding and linear CDF decoding run on the CPU of every device. The hyper decoder and the two entropy-model steps, which together account for 1.56 of the receiver's 5.21 kMAC/pixel (Table 4), do not participate in symbol decoding; they run afterward on the floating-point or NPU backend to produce the quantization step qq and the means μ0\mu_0 and μ1\mu_1 that dequantize the decoded symbols. Table 6 therefore verifies agreement of the integer CPU path across three platforms, and says nothing about whether the reconstruction agrees. When a bitstream encoded on one device is decoded on another, a numerical difference in qq or μ0\mu_0 perturbs y^0\hat{y}_0, propagates through the second step into μ1\mu_1 and y^1\hat{y}_1, and alters the reconstruction, and the paper does not report the pixel-level difference between decodes of the same bitstream on the three backends. The paper's wording is accurate where it says "entropy coding"; Figure 2 and the Section 1 description of the "compact entropy-control path" should make the boundary explicit, and the cross-device reconstruction difference should be measured.

  5. The authors withhold the source code, the trained PULSE, PULSE-S, PULSE-L, and perceptual checkpoints, the INT8 linear-CDF weights and CDF tables, and the logs of the evolution rounds. There is no supplementary material and no anonymous repository link. The central claims depend on single-thread timings, which are sensitive to the inference runtime and rANS implementation, and on cross-device determinism, neither of which a reviewer can check without code. Since the code will not fit into the rebuttal text, the authors should provide it through an anonymous repository (e.g., https://anonymous.4open.science) during the rebuttal period.

References:

[a] Wieckowski et al., "Towards a Live Software Decoder Implementation for the Upcoming Versatile Video Coding (VVC) Codec," ICIP 2020.

[b] Blard et al., "Overfitted Image Coding at Reduced Complexity," EUSIPCO 2024.

[c] Jacobellis and Yadwadkar, "Learned Compression for Compressed Learning," DCC 2025.

[d] Hojjat et al., "MCUCoder: Adaptive Bitrate Learned Video Compression for IoT Devices," DAGM GCPR 2025.

[e] Alakuijala et al., "JPEG XL Next-Generation Image Compression Architecture and Coding Tools," SPIE Applications of Digital Image Processing XLII, 2019.

Questions for authors
  1. What is the intended deployment scenario (delivery to client CPUs, capture on a device, or another), and how do PULSE and PULSE-S compare in BD-rate and measured single-thread decoding time with libjxl, dav1d (AVIF), an optimized VVC decoder [a], and an overfitted codec such as C3 or Cool-Chic on the same CPU?

  2. For the hardware-aware NAS row in Table 2 (1280 training runs): did the authors run the search, and what were its search space, sampling procedure, per-candidate training schedule, and complexity constraint? On which images were evolution candidates selected, and what is the seed-to-seed BD-rate variation of a fixed candidate?

  3. Because the means μi\mu_i and the quantization step qq are computed by the neural backend, outside the integer CPU path, what are the maximum absolute pixel difference and the PSNR between decodes of the same bitstream on the three platforms of Table 6, and between the encoder's reconstruction and each of them?

  4. What is HEIC's BD-rate with x265 in 4:4:4 at matched settings, in which color space is PSNR computed, and why was 4:4:4 worse?

  5. What are the character error rate of a text recognizer and the face verification accuracy on decoded ROI crops for PULSE (Perceptual) with and without the Stage II ROI losses, compared with MS-ILLM, and on which images and with which detector was Table 8 evaluated?

  6. Will the authors provide an anonymous repository with the inference code (including the INT8 linear-CDF path and rANS), checkpoints, and timing scripts during the rebuttal period?

15 comments

W1 & Q1: Deployment perspective and the missing operating point

We do not view deployment as a binary choice between image delivery and mobile capture. The more relevant question is whether a codec can remain practical across the increasingly heterogeneous compute available on modern devices.

Why neural codecs are attractive for deployment. CPUs remain the common denominator across devices; recent phones and PCs increasingly include NPUs, while high-end devices and clusters commonly provide GPUs. Conventional codec software runs primarily on CPUs and cannot readily reuse NPU/GPU compute. Fixed-function codec hardware, such as VPUs, can be highly efficient, but it is codec-specific, not portable across platforms, and difficult to update after deployment, especially on the decoder side. Neural codecs provide a complementary path because they run on programmable CPU, GPU, and NPU backends and can evolve through software. This also allows them to benefit directly from advances in general-purpose hardware and software runtimes; once deployed, future acceleration may not require redesigning and rolling out new fixed-function codec hardware.

The missing operating point addressed by PULSE. Existing practical neural codecs can achieve real-time performance on powerful accelerators, but often remain too expensive when only limited compute is available. PULSE is designed to span this compute range rather than target only high-end hardware. The single-thread CPU setting is therefore a stringent pressure test, not the only deployment target. CPU-only systems can exploit additional cores for lower latency, while NPU/GPU-equipped systems can accelerate the same codec further. The same bitstream can be decoded across these backends.

Perceptual optimization beyond distortion. Beyond hardware scalability, the neural formulation allows the compression objective itself to be adapted. PULSE is optimized not only for distortion but also for subjective quality and the preservation of faces and text. PICO [f] shows that perceptual optimization can reduce bitrate by a factor of 2–3 at comparable human visual quality; Tables 3 and 9 evaluate PULSE in this regime.

In short, PULSE targets a hardware-scalable practical neural codec: it remains usable when only CPUs are available, benefits from modern NPU/GPU acceleration when present, and retains the perceptual flexibility of end-to-end learned compression.

W1 & Q1 & Q4: Comparison with deployed and overfitted codecs

We reran the comparison at 1080p and measured single-thread CPU coding time under a common protocol. All deployed-codec comparisons in the main table use 4:4:4, PSNR is measured in RGB8 after decoding, and timing excludes image I/O. GPU-encoded and published-setting results are marked and discussed separately.

MethodEncoderEncoding timeDecoderDecoding timeBD-rate vs. HM
HMHM 16.258.8 sHM 16.2598 ms0.0%
Deployed Codecs
JPEGlibjpeg-turbo0.019 slibjpeg-turbo8.9 ms+288.3%
JPEG XL [e]libjxl0.98 slibjxl76 ms+77.7%
HEIC 4:4:4libheif + x2651.93 slibheif + libde265103 ms+19.1%
AVIF 4:4:4libaom8.83 slibdav1d98 ms-5.9%
VVC 4:4:4 [a]VTM 24.0406.17 sVVDeC 3.3140 ms-27.3%
Overfitted Codecs
C3–2310 s (GPU)–614 ms-24.1%
Cool-Chic 3.2 HOP [b]standard~19,500 s (GPU)–643 ms-22.8%
Cool-Chic 3.2 VLOP [b]standard~12,000 s (GPU)–131 ms-5.8%
Cool-Chic 3.2 VLOP [b]fast345 s (GPU)–131 ms+5.0%
Other Weak-Sender Systems
WaLLoC [c]–0.49 s–41.07 s+251.6%
MCUCoder [d]–0.11 s–67.13 s+534.7%
The Proposed
PULSE-S–0.92 s–77 ms+5.9%
PULSE–0.92 s–126 ms-2.9%
  • For WebP, libwebp does not support 4:4:4. Under 4:2:0, WebP reaches +78.4% BD-rate versus HM, with 400 ms encoding and 17.5 ms decoding. We therefore exclude it from the matched 4:4:4 table.
  • For VVC [a], the matched 4:4:4 result uses VTM because VVenC does not support 4:4:4. VVenC 4:2:0 encodes in 16.15 s but reaches only -15.7% versus HM on this set.
  • For HEIC 4:4:4, we originally tested libheif with psy-rd=psy-rdoq=1 enabled by default in the paper. This causes +6 Cb/Cr QP offsets to 4:4:4. Disabling it corrects the 4:4:4 result to +19.1%, compared with +25.2% for the original 4:2:0 setting.

(1) Production codecs.

  • Compared to JPEG XL [e] and HEIC, PULSE-S substantially outperforms both in PSNR rate–distortion, while encoding faster and decoding at comparable or lower latency.
  • Compared to AVIF and VVC, PULSE does not outperform them in PSNR BD-rate, but offers much lower encoding cost: 0.92 s, compared with 8.83 s for AVIF and 406.17 s for VVC. Encoding efficiency remains important for large-scale image distribution, such as social-media platforms and image galleries.
  • More importantly, PULSE supports perceptual optimization. Although this advantage is not captured by PSNR BD-rate, perceptual coding can reduce the required bitrate by a factor of 2–3 at comparable human visual quality, as also reported by PICO [f]. We report the corresponding perceptual comparisons in Tables 3 and 9.

(2) Overfitted codecs.

  • C3 and Cool-Chic [b] incur very high per-image GPU encoding costs, ranging from hundreds to tens of thousands of seconds. Even for image delivery, this cost is prohibitive. Under the official fast encoding setting, the BD-rate of Cool-Chic degrades to +5.0%, worse than PULSE.
  • Due to the highly sequential nature of their entropy models, the VLOP Cool-Chic decoder remains slightly slower than PULSE and substantially slower than PULSE-S.

(3) Other weak-sender systems.

  • WaLLoC [c] and MCUCoder [d] prioritize low sender complexity but incur substantially higher decoding costs and markedly worse PSNR rate–distortion than PULSE.
  • Their objectives and evaluation settings also differ: WaLLoC targets compressed-domain learning, while MCUCoder is trained for MS-SSIM and resizes inputs to 224×224. We therefore present these results as contextual rather than directly matched comparisons.

W2 & Q2: Agentic evolution and NAS comparison

(1) Details of the NAS baseline. We reran hardware-aware NAS rather than citing a literature value.

  • Starting from E1 (Figure 10 in the paper), we exhaustively searched N∈{1,2,3,4,5}N\in\{1,2,3,4,5\}, CR∈{1,2,4,8}C_R\in\{1,2,4,8\}, Y,CEM∈{32,64,…,384}Y,C_{EM}\in\{32,64,\ldots,384\}, Z∈{32,64,128,192}Z\in\{32,64,128,192\}, and CD∈{32,48,…,256}C_D\in\{32,48,\ldots,256\}.
  • Validity checks and the ±2.5% constraint around 2.5 kMAC/pixel retained 1,280 architectures, all of which were trained.
  • For the budget-matched 22-run NAS control, we uniformly sampled 22 architectures from the same set of 1,280 valid candidates.
  • Each hardware-aware NAS and agentic candidate was trained with SOAP for 300k steps: 200k at 10−310^{-3} and 100k at 10−410^{-4}. Both methods used Kodak PSNR BD-rate for selection, with E1 as the common anchor; CLIC validation and test were not used for selection.
  • One isolated training run requires approximately 6 H100-hours. The exhaustive NAS cost is therefore 7,680 serial-equivalent H100-hours and approximately 4,200 actual H100-hours with colocated jobs (e.g., four concurrent runs on one H100).

(2) Controlled ablations of human review, probes, search space, and search procedure. The new axes were discovered through probe-guided agentic evolution and were absent from the predefined NAS grid. We add the following controls, including a budget-matched 22-run NAS baseline, to separate the contributions of human review, probes, search efficiency, and search flexibility:

WorkflowHumanNew search axesProbes# Training RunsH100 GPU-HoursBD-rate
Full agentic evolution✓✓✓2278-10.2%
LLM + probes✗✓✓2278-8.7%
Agentic evolution✗✗✓2278-7.1%
LLM only✗✗✗2278-3.9%
Hardware-aware NAS✗✗✗2272-3.1%
Hardware-aware NAS✗✗✗1,280~4,200-7.5%
DARTS✗✗✗876-5.3%
  • Human review. Removing human review reduces the BD-rate improvement by 1.5 percentage points, from -10.2% to -8.7%. Nevertheless, the human-free, probe-guided LLM still outperforms exhaustive NAS with only 22 rather than 1,280 training runs.
  • Search efficiency. Under the same predefined axes and the same 22-run budget, agentic evolution reaches -7.1% using 78 H100-hours, compared with -3.1% using 72 H100-hours for hardware-aware NAS. With only 22 runs, it also approaches the exhaustive NAS result of -7.5% obtained from 1,280 runs.
  • Search flexibility. Allowing the LLM to discover and introduce new axes further improves the result from -7.1% to -8.7%. This ability to evolve the search space during search is a key advantage of agentic evolution over fixed-space NAS.
  • Probe contribution. Under the same fixed search space and 22-run budget, removing probes reduces the result from -7.1% to -3.9%, a 3.2 percentage-point reduction.
  • Efficient NAS. Following the DARTS protocol, we conducted four independent search runs and fully trained each resulting architecture, for eight runs in total. Although DARTS requires fewer runs, its supernet search is computationally expensive, resulting in a total cost comparable to agentic evolution (76 vs. 78 H100-hours). Under this comparable budget, agentic evolution achieves a larger BD-rate improvement (-10.2% vs. -5.3%).

A prospective human-only study using the same probes would be informative. However, a post-hoc rerun by the authors would be biased because the researchers already know the successful directions and newly discovered axes. A rigorous human-only control would require independent experts who are unaware of the original trajectory while controlling for expertise, time, and training budget. We therefore do not claim that human–LLM evolution outperforms expert-only design.

(3) Seed variation. To quantify seed variation, we trained every architecture evaluated throughout the evolution process with three seeds under the same protocol. Across all architectures, the average BD-rate standard deviation is 0.31%, and all gains used to promote candidates during agentic evolution exceed this one-standard-deviation variation.

(4) Evaluation set. For architecture selection, we use Kodak rather than the CLIC Professional validation set. The stronger performance on CLIC results from the high-resolution training stage applied after architecture search, as detailed in Table 7.

Reviewer_Wm5pOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a framework for quantization and architecture optimization to support on-device deployment of neural image compression. It combines a meta-prior with integer quantization to ensure cross-device consistency and uses probe-guided human–LLM collaboration to improve system performance. Improvements over existing methods are demonstrated in a single-threaded CPU setting.

Critical strengths

The paper provides a relatively systematic solution addressing both quantization and architecture optimization.

Critical weaknesses
  1. For architecture optimization, the effectiveness of individual components within the LLM-driven framework is insufficiently validated, and comparisons with existing approaches are limited. The current evidence therefore falls short of establishing a solid, novel technical contribution.

  2. Latency is systematically compared with existing learned image compression (LIC) methods only on a single CPU thread, whereas the comparison with NAS methods is conducted on an NPU. Although the title explicitly limits the scope, an approach targeting on-device deployment should aim to demonstrate benefits beyond a single-threaded CPU setting.

Questions for authors
  1. How does the framework compare with efficient NAS methods such as DARTS (Differentiable Architecture Search)? With an appropriately constructed supernet, such methods could require less training compute than the framework’s 22 training runs. How do these approaches compare on devices other than NPUs?
  2. Can rigorous controlled experiments establish the benefits of the LLM-driven framework’s design? For example, could a double-blind comparison demonstrate whether the heuristic representation probes lead to better outcomes?
  3. The quantization approach appears to combine existing techniques. Was this design also developed through collaboration with the LLM?
  4. Can the proposed human-LLM collaboration framework be extended beyond architecture optimization?
4 comments

W1 & W2 & Q1: Comparison with efficient NAS

As suggested, we evaluated DARTS using the same training and evaluation protocol as in Table 2 of the paper. Following the DARTS paper, we conducted four independent search runs and fully trained each resulting architecture. DARTS is substantially cheaper than exhaustive hardware-aware NAS but achieves a smaller BD-Rate reduction. Under a comparable compute budget, agentic evolution delivers the largest improvement.

MethodBD-Rate# Training RunsH100 GPU-HoursSingle-Thread CPU LatencyGPU Latency
Hardware-aware NAS (budget-matched)-3.1%227278 ms0.8 ms
Hardware-aware NAS (exhaustive)-7.5%1,280~4,20080 ms0.7 ms
DARTS-5.3%87675 ms0.9 ms
Agentic evolution-10.2%227877 ms0.7 ms
  • Notably, although DARTS requires fewer training runs than agentic evolution, their total GPU costs are comparable because its search stage requires training an entire supernet.
  • GPU latency is measured on an NVIDIA RTX 4090.

To compare the resulting architectures beyond the NPU setting, we report both single-thread CPU and GPU latency. All architectures have similar latency on both devices because they satisfy the same receiver-complexity constraint; agentic evolution therefore achieves the largest BD-Rate reduction without sacrificing inference efficiency.


W1 & Q2: Controlled experiments on agentic evolution

We compare four variants under the same training protocol and the same budget of 22 training runs. To avoid information leakage, we use fresh agents with no access to the original evolution trajectory or its outcomes.

WorkflowHumanNew search axesProbesBD-rate vs. E1
Full agentic evolution✓✓✓-10.2%
LLM + probes✗✓✓-8.7%
Agentic evolution✗✗✓-7.1%
LLM only✗✗✗-3.9%

Human review improves the result by 1.5 percentage points, from -8.7% to -10.2%. Under fixed search axes, removing probes reduces the gain from -7.1% to -3.9%, while allowing new-axis discovery improves it from -7.1% to -8.7%.


Q3: Scope of LLM involvement in quantization

No. Agentic evolution was used only for receiver-architecture search in Section 4. The quantization and entropy-coding design in Section 3 was developed independently, without agentic evolution. We will clarify this scope.


Q4: Extension beyond architecture optimization

Yes. Our exploratory studies on other compression problems, including inference-speed optimization and training-data filtering, suggest that agentic evolution has broader potential. We also observed that effective transfer benefits from combining human expertise with task-specific probes: experts identify promising directions, while LLMs are often effective at proposing informative probes. However, these experiments are still preliminary, and some were conducted after submission. We therefore present these observations as directions for future work rather than as evidence supporting the claims of this paper.

We sincerely appreciate your time and effort in reviewing our work and providing constructive feedback. As the rebuttal deadline approaches, we would be grateful if you could let us know whether our responses have adequately addressed your concerns or if any further clarification is needed. Thank you again for your time and consideration.

Thank you for the rebuttal and for sharing the technical extension exploration. The major weakness has been addressed, and I will raise my score. In the camera-ready version or the next version, I suggest that the authors consider:

  • include the comparison with DARTS in the main text
  • include a discussion of future work in the appendix.
  • revise the abstract to better distinguish the novel technical contribution from the achieved results.

Best, The Reviewer.

We sincerely appreciate your careful review and constructive feedback, which have been invaluable in improving our paper. We will address the points you raised in the camera-ready version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.