PULSE: Unlocking Practical Image Compression on Single-Thread CPU
Zhaoyang Jia, Tianyu Zhang, Zihan Zheng, Wenxuan Xie, Jiahao Li, Bin Li, Houqiang Li, Yan Lu
cs.CV
Quick verdict
PULSE decodes 1080p in 126 ms on one CPU thread at HM-comparable quality with bit-exact entropy coding; strong systems work, but the RD claim is benchmark-dependent.
Summary
PULSE is a learned image codec targeting ultra-low decoding complexity on CPUs. Its 5.2 kMAC/pixel receiver combines depthwise-conv blocks, a two-step checkerboard entropy model, and a synthesis path that renders pixels from a 1/4-resolution grid via AdaLN-modulated MLPs and PixelShuffle. For bit-exact entropy coding across devices, an INT8 linear projection maps hyperlatents directly to indices of a 64-row log-spaced Gaussian CDF table (1.6 ms per 1080p image on one CPU thread), while a post-trained bank of 64 EM-assigned factorized priors (meta prior) recovers rate lost to this simplification. Architecture design used a human-LLM 'agentic evolution' loop guided by PCA-redundancy and per-image headroom probes: seven rounds and 22 from-scratch training runs improved BD-rate by 10.2% over the initial model. On CLIC Professional validation, PULSE reaches -2.9% PSNR BD-rate vs HM-16.25 with 126 ms single-thread 1080p decoding; a perceptual variant with DINOv2-discriminator and ROI training reaches -22.5% FID BD-rate vs MS-ILLM. Limitations: asymmetric encoding (920 ms single-thread), HM-comparability inconsistent across test sets (+0.3% on CLIC test, +1.2% on Kodak), perceptual advantage metric-dependent (worse than MS-ILLM on LPIPS-Alex), and an inherently hard-to-replicate human-LLM search process.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 4 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A 5.2 kMAC/pixel end-to-end neural receiver decodes a 1080p image in 126 ms on a single CPU thread while achieving -2.9% PSNR BD-rate vs HM-16.25 on CLIC Professional validation.
- An integer linear projection from hyperlatents to CDF-table indices plus a 64-bank EM-trained meta prior yields bit-exact cross-platform entropy coding with ~0.04% BD-rate loss and 1.6 ms CPU time per 1080p image.
- A human-LLM agentic evolution loop with PCA-redundancy and headroom probes improves BD-rate by 10.2% over seven rounds using only 22 training runs, versus 1280 runs for hardware-aware NAS.
- Two-stage perceptual optimization (LPIPS-heavy pretraining, then preference optimization with a DINOv2 discriminator and ROI losses) makes PULSE-L competitive with MS-ILLM at -22.5% FID BD-rate and far lower receiver complexity.
Strengths
Genuinely new efficiency operating point, validated end-to-end
Receiver is 5.2 kMAC/pixel vs ~79 (EVC-S) and ~21 (Shallow NTC) in Table 3; measured full-pipeline latency includes entropy decoding: 126 ms single-thread 1080p on EPYC 9V84, 20 ms decode / 42 ms encode on a Snapdragon 8 Elite NPU phone (Fig. 1, Sec. 6.3), with cross-device bit-exactness verified (Table 6).
Well-controlled bit-exactness ablation
Table 1 isolates each entropy-coding design with both BD-rate and CPU time: full integerization 85.9 ms / +5.7% INT8, LUT scale decoder 0.5 ms / +29.7%, hyper scale decoder 28.1 ms / +3.2%, linear CDF decoder 1.6 ms / +2.4%, plus meta prior to +0.04% — a clean cost/quality ladder supporting the core design choice.
Transparent complexity and training accounting
Honest scoping of claims
Weaknesses
Several anchors are third-party reimplementations without reported cross-checks
HM-comparability headline is benchmark-dependent
Search evidence is a single trajectory with an uncontrolled NAS comparison
Perceptual 'competes with MS-ILLM' depends on metric selection
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How fast does PULSE decode images?
Per the paper, PULSE decodes a 1080p image in 126 ms on a single thread of an AMD EPYC 9V84 CPU (the smaller PULSE-S takes 77 ms). On a Redmi K80 Pro phone with a Snapdragon 8 Elite NPU, the full pipeline takes about 42 ms to encode and 20 ms to decode at 1080p. Encoding on a single CPU thread is much slower, about 920 ms, so the design favors decoding on constrained devices.
Is PULSE better than HEVC (HM) or VVC (VTM)?
On PSNR BD-rate over CLIC Professional validation, PULSE is 2.9% more efficient than HM-16.25 but clearly behind VTM-17.0 (which is 23.4% better than HM). On other sets such as CLIC test and Kodak, PULSE is roughly at parity with HM. Its selling point is not raw compression but decoder speed (5.2 kMAC/pixel, 126 ms single-thread 1080p decode) and bit-exact entropy coding across devices.
What does bit-exact entropy coding mean in PULSE and why does it matter?
Different hardware produces slightly different floating-point results, and in an entropy decoder a single changed probability-table lookup can desynchronize the whole bitstream. PULSE computes its CDF-table index with an integer-only linear projection that runs identically on any CPU (1.6 ms per 1080p image), so decoding is bit-identical on a phone NPU, a laptop CPU, and a GPU server; the paper reports zero mismatches across three devices.
Is the PULSE code and training data available?
The paper states code is available at github.com/microsoft/GenCodec/tree/main/PULSE; this review could not verify the repository contents. Appendices disclose training data (a 4.8M-image subset of SAM-1B plus Open Images for low-res training, DIV2K/Flickr2K/CLIC for high-res), full schedules, and hyperparameters for both the MSE and perceptual models.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (RD curves, architecture diagram, visual comparisons) are not analyzable from extracted text; RD judgments rely solely on reported BD-rate tables.
- The extraction includes main body and appendices with tables, but figure-adjacent text shows OCR garbling; some appendix details (e.g., exact evaluation image counts) may be missing.
- The claimed code repository was not verified (no network access during review).
- This is an arXiv preprint (2609.18602, Sep 2026) and has not, to this reviewer's knowledge, completed peer review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.