ProGIC: Progressive and Lightweight Generative Image Compression with Residual Vector Quantization

Hao Cao; Chengbin Liang; Wenqi Guo; Zhijin Qin; Jungong Han

cs.CVMar 3, 2026

ProGIC: Progressive and Lightweight Generative Image Compression with Residual Vector Quantization

Authors: Hao Cao, Chengbin Liang, Wenqi Guo, Zhijin Qin, Jungong Han

Organizations: Tsinghua University · State Key Laboratory of Space Network and Communications

Abstract

Recent advances in generative image compression (GIC) have delivered remarkable improvements in perceptual quality. However, many GICs rely on large-scale and rigid models, which severely constrain their utility for flexible transmission and practical deployment in low-bitrate scenarios. To address these issues, we propose Progressive Generative Image Compression (ProGIC), a compact codec built on residual vector quantization (RVQ). In RVQ, a sequence of vector quantizers encodes the residuals stage by stage, each with its own codebook. The resulting codewords sum to a coarse-to-fine reconstruction and a progressive bitstream, enabling previews from partial data. We pair this with a lightweight backbone based on depthwise-separable convolutions and small attention blocks, enabling practical deployment on both GPUs and CPU-only devices. Experimental results show that ProGIC attains comparable compression performance compared with previous methods. It achieves bitrate savings of up to 57.57% on DISTS and 58.83% on LPIPS compared to MS-ILLM on the Kodak dataset. Beyond perceptual quality, ProGIC enables progressive transmission for flexibility, and also delivers over 10 times faster encoding and decoding compared with MS-ILLM on GPUs for efficiency.

Figures & tables

Figure 1: BD-rate vs. Decoding Latency on the Kodak dataset measured with DISTS on one NVIDIA A100 GPU. The proposed ProGIC attains competitive BD-rate while substantially reducing decoding latency. Upper-left indicates better.

Figure 2: Conceptual illustration of the motivation behind ProGIC. The original image vector is approximated by a base vector plus a sequence of residual vectors, yielding progressively improved reconstructions.

Panel of Figure 3 — Figure 3: (a) Overview of the proposed ProGIC. Each down-/up-sampling stage consists of a stack of $M$ depthwise convolution blocks and a feed-forward network (FFN). The blocks in $g_{s}(\cdot)$ are modified with feature modulation, as described in Sec. 3.2 . (b) Depthwise convolution block. “Depth conv” denotes a depthwise convolution, while others are pointwise convolutions. (c) FFN architecture, where “Chunk-2” splits the tensor into two equal parts along the channel dimension.

Figure 4: Feature modulation in an FFN: at each progressive decoding stage, stage-specific scale and bias are applied to the features before the residual addition.

Figure 5: Rate-distortion performance on the Kodak, Tecnick, DIV2K, and CLIC2020-Professional datasets, evaluated with LPIPS and DISTS vs. BPP. Curves closer to the lower-left are better, indicating better quality at the same compression ratio. “OOM” denotes out-of-memory under the official evaluation environment.

Method	Enc.(ms)	Dec.(ms)	FLOPs(G)	Params(M)	BD-rate (LPIPS)
Method	Enc.(ms)	Dec.(ms)	FLOPs(G)	Params(M)	Kodak	Tecnick	DIV2K	CLIC 2020
VTM-23.10 (Intra) [ 50 ]	$>$ 9999	150.30	–	–	313.84%	295.19%	285.10%	498.64%
LIC-HPCM [ 36 ]	62.37	82.88	732.47	68.50	274.50%	305.42%	267.18%	745.04%
DCVC-RT (Intra) [ 26 ]	14.09	17.08	382.98	45.65	393.72%	329.44%	349.39%	584.41%
HiFiC [ 41 ]	526.51	1408.60	599.51	181.60	45.82%	68.66%	46.36%	86.45%
Control-GIC [ 34 ]	103.56	436.26	5816.37	130.36	33.36%	68.83%	73.77%	136.25%

Table 1: BD-rate and computational complexity on the Kodak dataset, evaluated on an NVIDIA A100 GPU. Best results are in bold . Second best are underlined . A dash (–) denotes unavailable results. “Enc./Dec.” indicates encoding/decoding time per image.

Figure 6: Visualization of reconstructed images from different methods on Kodak. Values denote DISTS / bpp. Lower DISTS indicates better perceptual quality. Lower bpp indicates higher compression.

Method	256 $\times$ 256		512 $\times$ 512
Method	Enc.	Dec.	Enc.	Dec.
VTM-23.10 (Intra) [ 50 ]	$>$ 10s	330	$>$ 10s	456
DCVC-RT (Intra) [ 26 ]	247	187	981	725
MS-ILLM [ 42 ]	121	368	507	1352
DiffEIC [ 37 ]	–	$>$ 10s	–	$>$ 10s
OSCAR [ 20 ]	805	2530	3429	9519

Table 2: Laptop AMD Ryzen 7840HS CPU runtimes (ms) for image encoding and decoding across different resolutions. Best results are in bold . Second-best are underlined .

CPU	256 $\times$ 256		512 $\times$ 512
CPU	Enc.	Dec.	Enc.	Dec.
Snapdragon 870	0.563	0.675	2.498	3.827
MTK Dimensity 8000	0.559	0.715	2.346	3.658

Table 3: Mobile phones CPU runtimes (s) for image encoding and decoding with ProGIC-s across different resolutions.

	BD-rate	Enc.(ms)	Dec.(ms)
MS-ILLM	0.0 %	165.38	147.79
Base + ProgDTD (PSNR)	487.18 %	39.31	52.46
Base + ProgDTD (LPIPS)	-10.28%	38.37	51.38
Base + RVQ (LPIPS)	-48.10%	6.23	9.33
+ Attention	-56.00%	7.70	10.61
+ Attention + Modulation	-57.57%	7.70	10.62

Table 4: Ablation study on the effect of different components and entropy coding. “Base” is the lightweight backbone, and “ProgDTD” is the baseline progressive quantization method.

Figure 7: R–D performance with different codebook numbers.

Figure 8: Effect of different weighting ratios $p$ for training. The top-right zoom highlights the low-bitrate region. The bottom-right zoom highlights the high-bitrate region.

Figure 9: Visualization of progressive image transmission with ProGIC. Reconstructions are shown across increasing stages, where the cumulative BPP increases and visual fidelity improves. Values beneath each image denote DISTS / BPP.

Figure 10: Reconstruction quality comparison across different resolutions, both compressed to a total of 1280 bytes. The left column shows the original image. The middle column shows the reconstructed image at 0.04 BPP and $512\times 512$ resolution. The right column shows the reconstructed image at 0.16 BPP and $256\times 256$ resolution. Resolution and compression FLOPs are indicated above and below each result, respectively.

Figure 11: Entropy of different codebook usages in ProGIC.

Figure 12: Rate-distortion curves with and without entropy coding.

Figure 13: Visualization by t-SNE of latent features. The top-left plot shows the distribution of the unquantized latent vector $\bm{y}$ . The subsequent plots illustrate the distributions of the base vector $\hat{\bm{y}}_{1}$ and the residuals from each quantization stage, respectively.

Panel of Figure 14 — Figure 14: In the event of a forest fire, ProGIC enables rapid response by transmitting images over a satellite short message link, assuming one transmission every 60 seconds. The leftmost column shows the original image. The first row illustrates progressive transmission: a usable preview is available immediately and is progressively refined as more bits arrive. In contrast, the second and third rows depict non-progressive schemes, which require the full bitstream before decoding. The second row uses a low BPP, yielding a fast but fixed-quality preview that cannot be improved later. The third row uses a high BPP, delivering higher fidelity but only after a long wait for the complete transmission, which delays situational awareness.

Figure 15: Different BPP ranges achieved by varying the number of codebooks and downsampling factors. The left panel shows the case of 16 codebooks with a downsampling factor of 32, with a minimal BPP increment of 0.009766. The right panel shows the case of 48 codebooks with a downsampling factor of 64, with a minimal BPP increment of 0.001953.

Method	BD-rate (DISTS)
Method	Kodak	Tecnick	DIV2K	CLIC
HiFiC [ 41 ]	90.08%	99.67%	100.76%	124.45%
Control-GIC [ 34 ]	34.18%	67.12%	62.09%	110.76%
MS-ILLM [ 42 ]	0.00%	0.00%	0.00%	0.00%
DiffEIC [ 37 ]	-33.79%	23.68%	15.78%	59.91%
OSCAR [ 20 ]	-50.63%	-4.76%	–	–

Table 5: Comparison of BD-rate on the Kodak, Tecnick, DIV2K, and CLIC datasets measured with DISTS. Best results are in bold . Second-best are underlined .

Method	512 $\times$ 768			1080 $\times$ 1920			1440 $\times$ 2560			2160 $\times$ 3840
Method	Enc.	Dec.	Mem.	Enc.	Dec.	Mem.	Enc.	Dec.	Mem.	Enc.	Dec.	Mem.
VTM-23.10 [ 50 ]	$>$ 10s	150.30	–	$>$ 10s	230.10	–	$>$ 10s	288.16	–	$>$ 10s	486.71	–
LIC-HPCM [ 36 ]	62.37	82.88	0.53	309.80	342.95	1.98	465.91	474.49	2.84	1121.92	1147.79	6.05
DCVC-RT [ 26 ]	14.09	17.08	0.34	76.68	59.87	1.04	135.87	102.95	1.73	259.86	197.83	3.63
HiFiC [ 41 ]	526.51	1408.60	1.14	2894.55	6909.92	2.97	5179.44	$>$ 10s	4.78	$>$ 10s	$>$ 10s	9.75
Control-GIC [ 34 ]	103.56	436.26	6.53	610.76	2186.30	69.99	1190.97	4361.94	34.93	–	$>$ 10s	–

Table 6: Comparison of GPU runtimes (ms) and memory (GB) for image encoding and decoding across different resolutions. Enc./Dec. denote encoding/decoding times. Mem. denotes memory usage. Best results are in bold . Second-best are underlined .

Figure 16: Rate-distortion performance on the Kodak, Tecnick, DIV2K, and CLIC2020-Professional datasets, evaluated with PSNR, MS-SSIM [ 53 ] , and CLIP-IQA [ 52 ] versus BPP. Curves closer to the lower left are better, indicating higher quality at the same compression ratio.

Figure 17: Visualization of reconstructed images from different methods on Tecnick. Values denote DISTS / BPP. Lower DISTS indicates better perceptual quality, and lower BPP indicates higher compression.

Figure 18: Visualization of reconstructed images from different methods on Tecnick. Values denote DISTS / BPP. Lower DISTS indicates better perceptual quality, and lower BPP indicates higher compression.

Figure 19: Visualization of reconstructed images from different methods on DIV2K. Values denote DISTS / BPP. Lower DISTS indicates better perceptual quality, and lower BPP indicates higher compression.

Figure 20: Visualization of reconstructed images from different methods on DIV2K. Values denote DISTS / BPP. Lower DISTS indicates better perceptual quality, and lower BPP indicates higher compression.

Figure 21: Visualization of reconstructed images from different methods on CLIC 2020. Values denote DISTS / BPP. Lower DISTS indicates better perceptual quality, and lower BPP indicates higher compression.

Figure 22: Visualization of reconstructed images from different methods on CLIC 2020. Values denote DISTS / BPP. Lower DISTS indicates better perceptual quality, and lower BPP indicates higher compression.

Generative Image Compression

Generative image compression uses compact codes and learned generative priors to reconstruct plausible images at very low bitrates.

See all 35 papers

Efficient Generative Image Compression with Quantized Latents Refinement
Oct 21, 2026Hao Cao, Yinhuan Huang, Wenqi Guo +2Generative Image Compression Learned Image Compression
Department of Electronic Engineering, Tsinghua University · Department of Automation, Tsinghua University · State Key Laboratory of Space Network and Communications +1
ProGVC: Progressive-based Generative Video Compression via Auto-Regressive Context Modeling
Mar 18, 2026Daowen Li, Ruixiao Dong, Kai Li +3Video Compression Generative Video Compression
Alibaba Group, China · Institute of Artificial Intelligence, Hefei Comprehensive National Science Center
Adaptive Fused Prior Transfer for Controllable Generative Image Compression
May 16, 2026Yifei Pei, Ying Liu, Nam LingLearned Image Compression Generative Image Compression
Department of Computer Science and Engineering, Santa Clara University, Santa Clara, CA 95053 USA

Oct 21, 2026J/K move · Enter open · S save

Efficient Generative Image Compression with Quantized Latents Refinement

Hao Cao, Yinhuan Huang, Wenqi Guo, Zhijin Qin+1

Department of Electronic Engineering, Tsinghua University · Department of Automation, Tsinghua University · State Key Laboratory of Space Network and Communications+1

Generative Image Compression Learned Image Compression Residual VQ

Open paper AskPDF arXiv

Generative image compression (GIC) delivers strong perceptual quality at low bitrates. However, GIC methods with strong perceptual quality often rely on pretrained diffusion models with high latency, while recent lightweight GIC methods reduce latency but deliver lower perceptual quality. We observe that, for lightweight GIC methods at low bitrates, coarse deterministic quantization can collapse nearby encoder latents corresponding to different fine details to the same point, making these details difficult to recover. Motivated by this, we propose GR-GIC, an efficient GIC that delivers competitive reconstruction performance with fast inference. For better performance, we introduce Guided Refinement (GR), which combines latent refinement guided by entropy context with adaptive perturbations for training regularization, jointly improving reconstruction at low bitrates. For fast inference, we introduce reorganized projected residual vector quantization (P-RVQ), which equivalently reorganizes residual updates into low-dimensional lookups to expose greater parallelism and reduce sequential computation. Experiments on Kodak show that for performance, compared with OneDC, GR-GIC achieves BD-rate savings of 23.19%, 37.71%, 22.30%, and 22.14% for LPIPS, DISTS, PSNR, and FID, respectively. For fast inference, the highly optimized implementation of GR-GIC encodes and decodes in 4.60 and 5.98 ms on an NVIDIA A100. Specifically, the reorganized P-RVQ implementation achieves a

3.14\times

average speedup over the original implementation under the same FP32 setting. On an NVIDIA Jetson Orin NX, GR-GIC achieves latency comparable to lightweight GIC methods.

Abstract

Generative Image Compression

Explore similar work

Efficient Generative Image Compression with Quantized Latents Refinement

ProGVC: Progressive-based Generative Video Compression via Auto-Regressive Context Modeling

Adaptive Fused Prior Transfer for Controllable Generative Image Compression