Channel-Level Mask Guidance for Slimmable Image Compression

Chen Chen; Chenggang Wu; Yongbiao Chen; Weidong Qiu; Zhengwei Qi; Haibing Guan

eess.IVOct 19, 2026

Channel-Level Mask Guidance for Slimmable Image Compression

Authors: Chen Chen, Chenggang Wu, Yongbiao Chen, Weidong Qiu, Zhengwei Qi, Haibing Guan

Organizations: School of Computer Science, Shanghai Jiao Tong University, Shanghai 200240, China

Abstract

Learned image compression (LIC) has achieved promising rate-distortion performance, but most methods lack scalability for diverse platforms with varying computational and storage constraints. Existing slimmable LIC methods achieve flexible complexity via consecutive-channel activation but suffer from two limitations: forced consecutive-channel selection degrades narrow-model performance at low bitrates, and the lack of stable variable-rate control limits adaptability to varying bandwidth. We propose a channel-level mask-guided slimmable image compression method. A learnable channel-wise mask with sparsity-promoting regularization enables non-contiguous top-w channel selection, letting narrow models use the most informative channels. After training, mask merging and channel reordering convert the model into a standard slimmable network via Mask-to-Slimmable (M2S) conversion, enabling efficient inference without architectural changes. For variable-rate control, we combine lambda- and width-conditioned modulation with quantization-step scaling and a progressive training pipeline stabilized by a freeze-then-fine-tune strategy. Experiments on Kodak across E2E, Hyperprior, and Checkerboard backbones show superior rate-distortion performance across widths and bitrates, with large gains for narrow models at low bitrates. At w=48, it achieves up to 19.22%, 24.11%, and 22.76% BD-rate savings over the same-width Slimmable baseline on E2E, Hyperprior, and CKBD, respectively. A single VBR model replaces 40 independent models, reducing storage by 14.8x to 15.9x. Code: https://github.com/sjtu-tcloud/MGSIC.

Figures & tables

Fig. 1: Three typical styles of Elastic LIC methods. (a) Down-Up Sampling Style. (b) Scalable Style. (c) Slimmable Style. Quantization and entropy coding are omitted. Element-wise addition could also be replaced by channel-wise concatenation.

Fig. 2: Overview of typical LIC model with context model. $g_{a}$ and $g_{s}$ consist of Convolution (CONV) / Transposed Convolution (TPCV) with Generalized Divisive Normalization (GDN) / Inverse GDN (IGDN), respectively. “AE/AD” denote arithmetic en-/de-coding. “Concat” denotes channel-wise tensor concatenation. “Q” denotes quantization. The numbers ①–④ denote the four positions for variable-rate methods summarized in Table I : ① feature maps, ② latent $y$ , ③ side information $z$ , and ④ the quantizer.

Fig. 3: Channel importance and its effect on reconstruction ( $\lambda=0.18,w=192$ ). (a) Channel-wise $\ell_{2}$ normalization of the latent $y$ , where each row represents 48 consecutive channels. (b) Joint reconstructions using only the top- $K$ channels ranked by $\ell_{2}$ norm, with $K\in\{5,25,50,100,192\}$ . PSNR and bpp are annotated on each reconstruction.

	Pos. a	Training settings b
AG-VAE [ 30 ]	②	Random training
DCVC [ 31 ]	②	Sequential training
CAE [ 27 ]	①④	Sequential training with bin size=1; random bin size
QVRF [ 25 ]	②	Train with $\lambda_{\max}$ , then sequential training with other $\lambda$ s
DCT-RES [ 32 ]	①	Random training, biased toward larger $\lambda$
MAE [ 26 ]	①	Sequential training with total loss of all $\lambda$ s

TABLE I: Summary of variable bitrate research.

Fig. 4: Overview of the proposed method. (Top) Training stage of the Mask model with learnable masks and sliced operators. (Bottom) Training and inference stage of the Slimmable model with variable-rate control, where $\lambda$ - and width-conditioned modulation networks generate combined variables via Hadamard product, applied to $g_{a}/g_{s}$ features and quantization step sizes for $y/z$ , and channel shuffling aligns $(\Phi,\Psi)$ across widths. The M2S conversion between the two stages is detailed in Fig. 8 . Operator configurations are in the supplementary material.

Fig. 5: Three channel selection strategies. (a) Traditional slimmable operator with a fixed prefix (

w_{\mathrm{in/out}}=3/2

). (b) Mask layer selecting top-

w

channels (

w=3

) via a learnable mask. (c) Sliced operator with index sets

\mathcal{I}_{\mathrm{in}}=[0,2,3]

\mathcal{I}_{\mathrm{out}}=[0,1]

, decoupling selection from execution. Gray borders in

y

indicate zero-padded data.

Fig. 6: Channel index propagation from

\hat{y}

h_{s}

and

g_{ep}

. The selected channels

\mathcal{I}_{y}

map to

\mathcal{I}_{\mathrm{out}}^{hs}=\bigcup_{i\in\mathcal{I}_{y}}\{2i,2i+1\}

h_{s}

and to

\mathcal{I}_{\mathrm{in}}^{ep}=\bigcup_{i\in\mathcal{I}_{y}}\{4i,4i+1,4i+2,4i+3\}

g_{ep}

. Unselected channels are zero-padded (gray).

Fig. 7: Checkerboard model for predicting

(\mu,\sigma)

\hat{y}

, and its slimmable adaptation. (top) Original two-pass prediction: Pass 1 uses only

\Psi

; Pass 2 uses

\Psi

and

\Phi

from decoded anchors. (bottom) Slimmable Model with contiguous active channels and channel shuffling. In Mask Model, channels may be non-contiguous and are handled by sliced operators with set

\mathcal{I}_{y}

Fig. 8: M2S conversion from the trained Mask model (top-left) to the Slimmable model (top-right) via ❶ mask merging and ❷ channel reordering. Mask merging absorbs the scaling factors of Mask $l$ into the preceding CONV $l$ (lower-left/middle). Channel reordering permutes the output channels of CONV $l$ , the input/output channels of GDN $l$ , and the input channels of the next CONV $l+1$ using the index set $\mathrm{Idx}_{l}$ derived from Mask $l$ (lower-right). The middle panel shows the intermediate state; $\mathrm{Idx}_{0}/\mathrm{Idx}_{1}$ are derived from Mask 0/1, and $C_{0}/C_{1}$ denote the output channel counts of the first two CONVs.

Variant	$\lambda$ -mod.	Q-step scaling	$w$ -mod.	Freeze	Stage
VBR	✓				2
VBR-Q	✓	✓			2
VBR-QF	✓	✓		✓	2
VBR-QW	✓	✓	✓		3
VBR-QFW	✓	✓	✓	✓	3

TABLE II: Comparison of variable-rate training variants.

Panel of Figure 11 — Fig. 9: RD performance comparison of Slimmable, Mask, and M2S on Kodak across five widths (48–192) for E2E, Hyperprior and CKBD backbones. M2S consistently outperforms both baselines at narrow widths (48, 72), while eliminating the regularization overhead of Mask at wide width (144, 192).

	Width	Mask (vs Slim)	M2S (vs Slim)	M2S (vs Mask)
E2E	48	-10.60	-19.22	-10.54
	72	-5.73	-14.44	-9.67
	96	+0.26	-8.33	-8.93
	144	+5.71	-2.77	-8.01
	192	+6.73	-1.29	-7.54
Hyper	48	-11.23	-24.11	-16.24

TABLE III: BD-rate savings (%) compared to the same-width Slimmable baseline on Kodak.

Fig. 10: PSNR gain (dB) of M2S over Slimmable across all widths and $\lambda$ values on Kodak. Each cell is obtained by cubic-spline interpolation of the RD curves at the minimum bpp for the given $\lambda$ and width.

Panel of Figure 14 — Fig. 11: RD performance comparison across five widths (48–192) on Kodak for (a) E2E, (b) Hyperprior, and (c) CKBD backbones. VBR-QW achieves the best performance on E2E and CKBD, while VBR-QFW remains the best on Hyperprior.

	Width	48	72	96	144	192
E2E	VBR	21.57	27.64	28.36	31.28	36.92
	VBR-Q	15.86	16.58	14.75	14.18	16.47
	VBR-QW	4.85	7.30	6.92	7.82	9.68
	VBR-QF	10.79	17.61	16.51	14.54	16.17
	VBR-QFW	10.37	17.36	16.70	14.53	14.84
Hyper	VBR	30.25	24.79	20.43	18.59	19.00

TABLE IV: BD-rate increase (%) over the M2S upper bound on Kodak. Best results per backbone are in bold.

		Params
Model	Width	$g_{a}$	$g_{s}$	$h_{a}$	$h_{s}$	FEM	$g_{cm}$	$g_{ep}$	ModNet	Total	Total (8 $\lambda$ s)
	48	183.65 K	183.60 K	136.08 K	255.98 K	2.93 K	115.30 K	78.77 K		956.36 K
	72	410.26 K	410.19 K	306.07 K	575.78 K	4.39 K	259.34 K	176.90 K		2.14 M
	96	726.72 K	726.63 K	544.03 K	1.02 M	5.86 K	460.99 K	314.21 K		3.80 M
	144	1.63 M	1.63 M	1.224 M	2.30 M	8.78 K	1.037 M	706.32 K		8.54 M
CKBD	192	2.89 M	2.89 M	2.18 M	4.09 M	11.71 K	1.84 M	1.26 M	0	15.16 M	244.80 M

TABLE V: Parameter comparison of the CKBD backbone across different widths. The CKBD row reports single-

\lambda

single-width models; the M2S row reports one model covering all five widths at a fixed

\lambda

; the VBR variants cover 5 widths

\times

\lambda

values with a single model. “Total (8

\lambda

s)” reports the total storage required if all eight

\lambda

values were deployed, i.e.,

8\times

the per-model total for single-

\lambda

models, and a single model’s total for the VBR variants.

Width	$g_{a}$	$h_{a}$	$h_{s}$	$g_{s}$	$g_{cm}$	$g_{ep}$	Enc ent_z	Enc ent_y	Dec ent_z	Dec ent_y	CDF (startup)
48	1.52	0.52	0.58	1.94	0.31	0.65	1.31	20.80	1.55	24.77	10.57–10.74
72	2.28	0.54	0.59	2.89	0.31	0.68	1.79	23.65	2.18	29.50	14.66–15.59
96	2.71	0.65	0.71	3.20	0.40	0.83	2.63	26.94	3.20	34.51	20.74–22.72
144	5.36	0.54	0.68	5.50	0.34	0.66	6.14	33.14	7.32	43.81	38.61–58.60
192	6.84	0.54	0.76	6.94	0.48	0.67	14.04	40.10	17.28	54.37	106.90–290.51

TABLE VI: Per-module latency (ms) of the CKBD (VBR-QW). “Enc/Dec ent_z” and “Enc/Dec ent_y” denote the encoding/decoding latency of the factorized entropy model for

z

and the Gaussian conditional entropy model for

y

, respectively. The last column is the one-time CDF construction cost, paid only when

\lambda/w

changes.

Fig. 12: Rate-distortion comparison on Kodak. E2E/Hyper/CKBD-192 denote variable-rate models at width 192. For CBANet, ELFIC, and EVC, the highest-quality RD curves among available models are used for reference.

	Ours
	E2E	Hyper	CKBD	ELFIC (Low Rate)	CBANet-HP	EVC (Scale_SL)
Params (M)	6.34	12.68	15.84	66.99	6.39	17.47
$N_{\lambda}$	8 (Interp.)			4 (Interp.)	4	4 (Interp.)
Rel. Cplx. (%)	25/37.5/50/75/100			7.66/35.82/74.06/100	25/50/100	25/50/75/100
Enc. (ms)	10/14/18/27/36	19/24/28/40/51	28/32/37/51/71	166	36	42/45/50/55
Dec. (ms)	14/20/25/38/50	23/31/36/51/66	32/38/45/62/86	175/193/199/202	54/56/58	32

TABLE VII: Parameter count and en-/de-coding latency on Kodak. Rel. Cplx. (%) is method-specific:

w/192\times 100\%

(ours), the decoder-computation ratio (ELFIC, CBANet), and

N_{\mathrm{enc}}/4\times 100\%

(EVC).

N_{\lambda}

is the number of default variable-rate configurations; “Interp.”: extra RD points via interpolation, not applicable to CBANet (independent bitrate-control modules).

Codec	BPP	PSNR (dB)	Enc. (ms)	Dec. (ms)
JPEG	0.17–2.12	21.48–37.33	2.22	2.28
JPEG2000	0.12–2.00	26.30–38.69	230.19	237.77
BPG	0.10–2.12	27.14–42.12	470.27	168.11
VTM-intra	0.11–2.11	28.43–41.17	140985.54	269.41

TABLE VIII: Summary of traditional codecs on Kodak. BPP and PSNR ranges are computed over all evaluated quality points. Enc./Dec. denote the average en-/de-coding time.

Channel Pruning

See all 10 papers

MixCompress: Mixture of Experts for Variable Rate Learned Image Compression
Jul 15, 2026Calvin-Khang Ta, Praneet Singh, Tong Shao +1Learned Image Compression Rate-Distortion Optimization
Dolby Laboratories, USA
Efficient Learned Image Compression without Entropy Coding
May 22, 2026Hao Cao, Wenqi Guo, Zhijin Qin +1Learned Image Compression Vector Quantization
Department of Electronic Engineering, Tsinghua University · Department of Automation, Tsinghua University · State Key Laboratory of Space Network and Communications +1
Every Packet Counts: Dispersing Information for Loss-Resilient Learned Image Compression
Aug 11, 2026Yuhang Wei, Chuqin Zhou, Yibo Shi +2Learned Image Compression
Shanghai Jiao Tong University · Huawei Technologies Ltd

Jul 15, 2026J/K move · Enter open · S save

MixCompress: Mixture of Experts for Variable Rate Learned Image Compression

Calvin-Khang Ta, Praneet Singh, Tong Shao, Peng Yin

Dolby Laboratories, USA

Learned Image Compression Rate-Distortion Optimization Mixture-of-Experts Models

Open paper AskPDF arXiv

Learned image compression (LIC) is bottlenecked by the need to store independent models for each rate-distortion operating point. Existing variable bit-rate (VBR) methods aim to reduce this overhead via dense parameter modulation, but forcing a shared backbone to approximate divergent mappings causes severe feature entanglement. Specifically, low-rate smoothing gradients inherently conflict with the preservation of high-frequency textural details, leading to sub-optimal performance. To resolve this, we propose MixCompress, a unified VBR framework based on sparse structural specialization. While sparsely gated Mixture-of-Experts (MoE) routing successfully mitigates gradient conflict, it operates on a fixed computational budget. To address the increased representational demands of higher bit-rates we introduce a Mixture-of-Depths (MoD) extension to dynamically scale model capacity. Combined with Conditional Auxiliary Transforms (CAT) for dynamic sub-band energy modulation, our hierarchical framework effectively dynamically scales capacity. Extensive evaluations demonstrate that MixCompress not only matches individually optimized single-rate baselines but can even surpass them, establishing a new Pareto frontier for computationally efficient image coding.

Abstract

Channel Pruning

Explore similar work

MixCompress: Mixture of Experts for Variable Rate Learned Image Compression

Efficient Learned Image Compression without Entropy Coding

Every Packet Counts: Dispersing Information for Loss-Resilient Learned Image Compression