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.CVSep 16, 2026

PULSE: Unlocking Practical Image Compression on Single-Thread CPU

Authors: Zhaoyang Jia, Tianyu Zhang, Zihan Zheng, Wenxuan Xie, Jiahao Li, Bin Li, Houqiang Li, Yan Lu

Organizations: University of Science and Technology of China · Microsoft Research Asia · Independent Researcher

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 5.2 kMAC/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 126 ms on a single CPU thread while achieving compression performance comparable to HM. After perceptual optimization, PULSE competes with larger perceptual codecs like MS-ILLM. Codes are at https://github.com/microsoft/GenCodec/tree/main/PULSE

Figures & tables

Figure 1: Pulse targets practical image compression on resource-constrained hardware. Its 5.2 kMAC/pixel receiver decodes a 1080p image in 126 ms on a single thread of an AMD EPYC 9V84, while enabling bit-exact entropy coding across devices like REDMI K80 Pro. On the CLIC Professional validation set, Pulse matches HM and its perceptual variant competes with larger codecs such as MS-ILLM ( Muckley et al., 2023 ) . Latency is measured for the full decoding pipeline, including network inference and entropy decoding. Circle radius denotes the decoding complexity.

Figure 2: PULSE architecture. It consists of an encoder at $1/16$ resolution, a decoupled linear CDF index decoder with a meta prior, a two-step entropy model, and a decoder with an AdaLN-modulated pixel renderer. AE and AD denote arithmetic encoding and decoding, respectively.

Figure 3: Bit-exact entropy coding path. Meta Prior transmits $\hat{\zeta}$ to select CDFs for $\hat{z}$ with different content complexities. An integer linear projection maps $\hat{z}$ to CDF indices for entropy coding of $y_{q}$ .

Method	CLIC test set		LC subset		Single-thread CPU Time
Method	Float	INT8	Float	INT8	$\hat{z}\to\sigma$	$\hat{z}\to$ CDF
Integerize full entropy model	1.5%	5.7%	3.1%	7.8%	85.9 ms	86.0 ms
+ LUT scale decoder ( Pärnamaa et al., 2026 )	29.7%	29.7%	59.6%	59.6%	0.5 ms	0.5 ms
+ Hyper scale decoder ( Ascenso et al., 2023 )	–0.1%	3.2%	2.1%	6.4%	28.1 ms	29.2 ms
+ Linear scale decoder	4.5%	4.5%	11.8%	12.0%	1.2 ms	4.1 ms
+ Linear CDF decoder	2.4%	2.4%	9.8%	9.9%	N/A	1.6 ms

Table 1: Ablation of bit-exact entropy-coding designs. BD-rate is measured on the CLIC test set and a subset of images with low content complexity (LC subset). Latency is measured for a 1080p image on one AMD EPYC 9V84 CPU thread.

Figure 4: Agentic evolution , illustrated for round E4 $\rightarrow$ E5 in the complete trajectory in Figure 10 . Layer-wise probes measure redundancy and headroom . A human researcher and an LLM agent interpret the evidence to propose fixed-budget reallocations, and validate them by training.

Figure 5: Visual comparison. Pulse achieves comparable or better visual quality after perceptual and ROI optimization. More results are in Appendix C.4 .

Figure 6: Rate–distortion curves. CLIC Professional validation results for MSE-optimized codecs (left) and perceptually optimized codecs (right). More results are in Appendix C.3 .

Table 2: Ablation studies. BD-Rate (

\downarrow

) is measured in terms of PSNR and FID for the MSE- and perceptual- optimized versions, respectively. Decoder speed is measured on 1080p images on an NPU.

Method	VBR	BE	Target	Encoding			Decoding			BD-rate $\downarrow$
Method	VBR	BE	Target	Params	kMAC/px	Time	Params	kMAC/px	Time	on PSNR	on FID
HEIC	$\checkmark$	$\checkmark$	MSE	libheif + x265		1.4 s	libheif + libde265		80 ms	+25.2%	+2680%
HM-16.25	$\checkmark$	$\checkmark$	MSE	HM Encoder		8.8 s	HM Decoder		98 ms	0.0%	+2450%
VTM-17.0	$\checkmark$	$\checkmark$	MSE	VTM Encoder		249 s	VTM Decoder		189 ms	-23.4%	+2660%
JPEG-AI	$\times$	$\checkmark$	MSE	7.1 M	56	2.6 s	4.4 M	8.1	643 ms	+37.4%	+1720%
MS-Hyperprior	$\times$	$\times$	MSE	14 M	112	2.1 s	12 M	109	2.4 s	-1.7%	+3830%

Table 3: Complexity and rate–distortion performance. Time measurements use

1088\times 1920

inputs on a single-thread CPU. BD-rate is evaluated on the CLIC Professional validation set. VBR denotes variable bitrates within one model, and BE denotes bit-exact entropy coding across platforms.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Sender	Receiver	PULSE		PULSE-S		PULSE-L
Component	Sender	Receiver	kMAC/px	Param.	kMAC/px	Param.	kMAC/px	Param.
Analysis transform	$\checkmark$	$\times$	43.22	12.1M	43.08	11.1M	43.21	12.1M
Hyper encoder	$\checkmark$	$\times$	0.45	0.30M	0.31	0.19M	0.93	0.66M
Hyper decoder	$\checkmark$	$\checkmark$	0.24	0.15M	0.16	0.09M	0.44	0.32M
Entropy model step-1	$\checkmark$	$\checkmark$	0.58	0.17M	0.48	0.13M	1.35	0.41M
Entropy model step-2	$\checkmark$	$\checkmark$	0.74	0.21M	0.61	0.16M	1.67	0.50M

Appendix

Table 4: Component accounting for Pulse .

Model	Pipeline	Neural inference	Others	Total
Pulse -s	Encoding	913 ms	7 ms	920 ms
Pulse -s	Decoding	70 ms	7 ms	77 ms
Pulse	Encoding	913 ms	7 ms	920 ms
Pulse	Decoding	119 ms	7 ms	126 ms
Pulse -L	Encoding	963 ms	9 ms	972 ms
Pulse -L	Decoding	411 ms	9 ms	420 ms

Appendix

Table 5: Single-thread CPU end-to-end latency of PULSE. Others includes non-neural processing, such as entropy coding and CDF indexing.

Environment	Platform	Neural backend	CPU backend	Mismatch
Phone	REDMI K80 Pro	NPU: Snapdragon 8 Elite	Snapdragon 8 Elite	0
Laptop	HP OmniBook Ultra Flip 14	CPU: Intel Core Ultra 9 288V	Intel Core Ultra 9 288V	0
Server	GPU server	GPU: NVIDIA A100	AMD EPYC 9V84	0

Appendix

Table 6: Cross-device entropy coding bit-exact validation. Mismatch counts both CDF-index and decoded-symbol mismatches.

Model	Stage	Crop	Steps	Batch	$\lambda$ range	Learning-rate schedule
MSE	LR training	$512$	$500$ k	8	$24$ – $672$	$10^{-3}$ (200k) $\to$ $10^{-4}$ (200k) $\to$ $10^{-5}$ (100k)
MSE	HR training	$1024$	$50$ k	8	$24$ – $672$	$10^{-4}$ (40k) $\to$ $10^{-5}$ (10k)
Perceptual	Stage I LR	$512$	$300$ k	16	$0.5$ – $12.5$	$10^{-3}$ (200k) $\to$ $10^{-4}$ (100k)
	Stage I HR	$1024$	$50$ k	16	$0.5$ – $12.5$	$10^{-4}$ (40k) $\to$ $10^{-5}$ (10k)
	Stage II	$1024$	$20$ k	32	$3$ – $25$	$10^{-4}$ (10k) $\to$ $10^{-5}$ (10k)

Appendix

Table 7: Training schedules.

Training variant	PSNR	LPIPS-VGG	FID	Face PSNR	Text PSNR
Stage I w/o high LPIPS weight	-0.4%	+6.0%	+5.7%	+6.1%	+1.6%
Stage II w/o GAN Loss	–3.6%	–6.4%	+49.8%	–4.1%	–4.5%
Stage II w/o ROI	+0.2%	–2.4%	–3.7%	+37.7%	+20.8%
Proposed perceptual optimization	0.0%	0.0%	0.0%	0.0%	0.0%

Appendix

Table 8: Ablation study on perceptual training designs. BD-Rate is measured.

Figure 7: Rate–distortion curves on the 300-image CLIC test subset released by PICO.

Method	PSNR	LPIPS-Alex	LPIPS-VGG	DISTS	FID
MRIC	-1.2%	+9.7%	+58.5%	+51.1%	+98.4%
MS-ILLM	0%	0%	0%	0%	0%
CoD-Lite	+160%	+109.4%	-39.7%	-1.3%	+9.4%
PICO	+18.2%	+50.7%	-30.6%	-52.3%	-6%
PULSE (Perceptual)	+2.2%	+31.1%	-38.2%	-29.3%	+12.7%
PULSE-L (Perceptual)	-7.1%	+8.9%	-48.9%	-45.2%	-18.2%

Appendix

Table 9: BD-rate comparison on PICO-matched 300-image CLIC test subset. Anchor: MS-ILLM.

Figure 8: Rate-distortion curves on more benchmarks for MSE-optimized codecs.

Figure 9: Rate-distortion curves on more benchmarks for perceptual-optimized codecs.

Figure 10: Full agentic evolution process and final parameters at different scales.

Transition	Main change	Motivation
E1 $\rightarrow$ E2	Move rendering to a $1/4$ -resolution grid	relieve the narrow full-resolution renderer
E2 $\rightarrow$ E3	Replace four narrow blocks with one wider block	concentrate low-resolution capacity
E3 $\rightarrow$ E4	Increase main and hyper latent capacity	address pressure in shared representations
E4 $\rightarrow$ E5	Introduce grouped pointwise convolution	reclaim entropy-model computation
E5 $\rightarrow$ E6	Reallocate internal width to the main latent	strengthen shared latent capacity
E6 $\rightarrow$ E7	Reduce FFN channels and increase decoder depth	improve nonlinear decoder processing

Appendix

Table 10: Illustrations on each evolution.

Method	Kodak	Tecnick	DIV2K	CLIC pro val	CLIC test
HEIC	+14.3%	+28.4%	+21.1%	+25.2%	+16.9%
HM-16.25	0%	0%	0%	0%	0%
VTM-17.0	-21.3%	-23.8%	-22.8%	-23.4%	-23%
JPEG-AI	+54.3%	+41.2%	+33%	+37.4%	+33.2%
MS-Hyperprior	-2.2%	-4.6%	-3.3%	-1.7%	-2%
Shallow NTC	+6.3%	+9.9%	+9.3%	+13%	+11.9%

Appendix

Table 11: BD-Rate comparison on PSNR. Anchor: HM-16.25.

Method	Kodak	Tecnick	DIV2K	CLIC pro val	CLIC test
HEIC	+890%	+820%	+900%	+1420%	+1330%
HM-16.25	+840%	+800%	+880%	+1340%	+1450%
VTM-17.0	+740%	+690%	+790%	+1140%	+1250%
JPEG-AI	+830%	+700%	+760%	+1130%	+1120%
MS-Hyperprior	+1010%	+920%	+1000%	+1630%	+1640%
Shallow NTC	+1530%	+1300%	+1540%	+2410%	+2360%

Appendix

Table 12: BD-Rate comparison on LPIPS-Alex. Anchor: MS-ILLM.

Method	Kodak	Tecnick	DIV2K	CLIC pro val	CLIC test
HEIC	+330%	+340%	+290%	+520%	+530%
HM-16.25	+300%	+330%	+270%	+490%	+560%
VTM-17.0	+260%	+270%	+230%	+480%	+540%
JPEG-AI	+270%	+240%	+200%	+360%	+370%
MS-Hyperprior	+300%	+350%	+290%	+570%	+580%
Shallow NTC	+460%	+540%	+450%	+950%	+1000%

Appendix

Table 13: BD-Rate comparison on LPIPS-VGG. Anchor: MS-ILLM.

Method	Kodak	Tecnick	DIV2K	CLIC pro val	CLIC test
HEIC	+550%	+590%	+570%	+1070%	+920%
HM-16.25	+480%	+560%	+500%	+960%	+960%
VTM-17.0	+570%	+630%	+610%	+1160%	+1130%
JPEG-AI	+560%	+550%	+540%	+910%	+860%
MS-Hyperprior	+680%	+890%	+790%	+1500%	+1410%
Shallow NTC	+980%	+1090%	+1070%	+2090%	+1920%

Appendix

Table 14: BD-Rate comparison on DISTS. Anchor: MS-ILLM.

Method	Kodak	Tecnick	DIV2K	CLIC pro val	CLIC test
HEIC	+770%	+1770%	+1940%	+2680%	+4820%
HM-16.25	+630%	+1630%	+1700%	+2450%	+4890%
VTM-17.0	+730%	+1480%	+1700%	+2660%	+3820%
JPEG-AI	+530%	+1140%	+1130%	+1720%	+2540%
MS-Hyperprior	+800%	+2060%	+2120%	+3830%	+4980%
Shallow NTC	+1010%	+2800%	+2990%	+5590%	+8750%

Appendix

Table 15: BD-Rate comparison on FID. Anchor: MS-ILLM.

Panel of Figure 11 — Figure 11: More visual examples on text contents.

Panel of Figure 12 — Figure 12: More visual examples on natural contents.

Learned Image Compression

Learned image compression uses neural encoders, quantization, and entropy models to reduce image bitrates while balancing fidelity, perceptual quality, and task utility.

See all 120 papers

What Matters in Practical Learned Image Compression
May 6, 2026Kedar Tatwawadi, Parisa Rahimzadeh, Zhanghao Sun +5Learned Image Compression Rate-Distortion Optimization
Apple
ARCHE: Autoregressive Residual Compression with Hyperprior and Excitation
Mar 10, 2026Sofia Iliopoulou, Dimitris Ampeliotis, Athanassios SkodrasLearned Image Compression Autoregressive Modeling
Dept. of Electrical and Computer Engineering, University of Patras, Patras, Greece · Dept. of Digital Media and Communication, Ionian University, Argostoli, Greece
PIC: Revisiting INR for Image Coding with Fast Encoding and Sub-Millisecond Decoding
Sep 8, 2026Xiang Liu, Jinxiang Wang, Bin Chen +5Efficient Neural Network Inference Learned Image Compression
Tsinghua University, China · Harbin Institute of Technology, Shenzhen, China · Peng Cheng Laboratory, China +1

May 6, 2026J/K move · Enter open · S save

What Matters in Practical Learned Image Compression

Kedar Tatwawadi, Parisa Rahimzadeh, Zhanghao Sun, Zhiqi Chen+4

Apple

Learned Image Compression Rate-Distortion Optimization

Open paper AskPDF arXiv

One of the major differentiators unlocked by learned codecs relative to their hard-coded traditional counterparts is their ability to be optimized directly to appeal to the human visual system. Despite this potential, a perceptual yet practical image codec is yet to be proposed. In this work, we aim to close this gap. We conduct a comprehensive study of the key modeling choices that govern the design of a practical learned image codec, jointly optimized for perceptual quality and runtime -- including within the ablations several novel techniques. We then perform performance-aware neural architecture search over millions of backbone configurations to identify models that achieve the target on-device runtime while maximizing compression performance as captured by perceptual metrics. We combine the various optimizations to construct a new codec that achieves a significantly improved tradeoff between speed and perceptual quality. Based on rigorous subjective user studies, it provides 2.3-3x bitrate savings against AV1, AV2, VVC, ECM and JPEG-AI, and 20-40% bitrate savings against the best learned codec alternatives. At the same time, on an iPhone 17 Pro Max, it encodes 12MP images as fast as 230ms, and decodes them in 150ms -- faster than most top ML-based codecs run on a V100 GPU.

Abstract

Learned Image Compression

Explore similar work

What Matters in Practical Learned Image Compression

ARCHE: Autoregressive Residual Compression with Hyperprior and Excitation

PIC: Revisiting INR for Image Coding with Fast Encoding and Sub-Millisecond Decoding