Dataset Color Quantization: A Training-Oriented Framework for Dataset-Level Compression

Chenyue Yu; Lingao Xiao; Jinhong Deng; Ivor W. Tsang; Yang He

cs.CVFeb 24, 2026

Dataset Color Quantization: A Training-Oriented Framework for Dataset-Level Compression

Authors: Chenyue Yu, Lingao Xiao, Jinhong Deng, Ivor W. Tsang, Yang He

Organizations: National University of Singapore · CFAR, Agency for Science, Technology and Research, Singapore · IHPC, Agency for Science, Technology and Research, Singapore · University of Electronic Science and Technology of China (UESTC) ·

Abstract

Large-scale image datasets are fundamental to deep learning, but their high storage demands pose challenges for deployment in resource-constrained environments. While existing approaches reduce dataset size by discarding samples, they often ignore the significant redundancy within each image -- particularly in the color space. To address this, we propose Dataset Color Quantization (DCQ), a unified framework that compresses visual datasets by reducing color-space redundancy while preserving information crucial for model training. DCQ achieves this by enforcing consistent palette representations across similar images, selectively retaining semantically important colors guided by model perception, and maintaining structural details necessary for effective feature learning. Extensive experiments across CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K show that DCQ significantly improves training performance under aggressive compression, offering a scalable and robust solution for dataset-level storage reduction.

Figures & tables

Panel of Figure 1 — Figure 1: Visualization of different Color Quantization algorithms on Tiny-ImageNet. Images are quantized into 4 colors, which are 2 color bits. (a) Original images. (b) Color quantization is performed through K-Means clustering to obtain representative color palettes for each image, wasting bits on backgrounds. (c) Independent and representative color palettes obtained by ColorCNN, which have abrupt textural discontinuities. (d) Our DCQ assigns more colors to foregrounds and has less textural discontinuity.

Panel of Figure 2 — Figure 2: Comparison of the original image and feature maps extracted from a ResNet-18 trained on CIFAR-10. Thermal color maps visualize activation strength from black to white, reflecting learned feature hierarchy.

Figure 3: The pipeline of our dataset color quantization framework. First, we apply K-means clustering to group images based on their features extracted from a pre-trained model. Next, within each cluster, we perform K-means on the color palettes of individual images to generate a shared color palette for all images in the same cluster. The generated palettes and their corresponding indices are stored for later use. During training, we retrieve the stored indices and palettes to reconstruct quantized images, which are then used to train a neural network.

Panel of Figure 4 — Figure 4: Visualization of four clusters obtained by applying ResNet-18 first-block features to partition CIFAR-10 into 20 clusters.

Panel of Figure 5 — Figure 5: Comparison between prior color quantization methods and our approach, evaluated with ResNet-18. Detailed results are provided in Appendix B . Unlike the original paper, which quantized the test set while training on the original train set, we quantize the entire train-set and keep the test-set unchanged.

	CIFAR-10					CIFAR-100					ImageNet-1K
Color Bits	5	4	3	2	1	5	4	3	2	1	5	4	3	2	1
Colors/Image	32	16	8	4	2	32	16	8	4	2	32	16	8	4	2
Prune Ratio	80%	83%	87.5%	92%	96%	80%	83%	87.5%	92%	96%	80%	83%	87.5%	92%	96%
Random	88.15	84.38	80.15	77.04	70.08	57.36	50.96	42.19	39.71	36.68	62.54	58.18	53.34	50.32	20.28
Entropy	79.08	74.59	67.58	64.51	57.06	41.83	35.45	29.77	28.96	22.16	55.80	44.59	42.04	31.04	17.06
EL2N	70.32	67.06	25.99	21.31	19.85	15.51	12.51	10.36	8.36	7.51	31.22	21.28	13.99	11.69	9.81

Table 1: Comparison of dataset pruning algorithms and our dataset color quantization algorithm on CIFAR-10, CIFAR-100 with ResNet-18, and ImageNet-1K with ResNet-34.

Table 2: Ablation Studies of Different Strategies.

	ResNet-34, CIFAR-10					ResNet-50, CIFAR-10
	80%	83%	87.5%	92%	96%	80%	83%	87.5%	92%	96%
Random	86.13	83.48	76.19	74.44	65.08	87.03	83.18	79.19	75.44	68.08
EL2N	70.21	66.36	24.69	23.11	22.19	69.71	65.96	24.59	23.31	21.89
AUM	58.34	49.59	29.98	21.60	20.35	57.84	49.09	29.58	25.60	21.35
CCS AUM	89.34	87.92	87.52	74.01	71.42	89.44	88.11	86.52	74.31	71.02
TDDS	89.58	88.15	86.92	73.05	69.48	90.68	88.55	87.62	74.01	70.46

Table 3: Comparison of dataset pruning algorithms and our algorithm on CIFAR-10 with ResNet-34 and ResNet-50.

Panel of Figure 6 — Figure 6: Illustration of the test accuracy of our method using ResNet-18 on CIFAR-10.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

	CIFAR-10						CIFAR-100						Tiny-ImageNet
	Color Bits						Color Bits						Color Bits
	1	2	3	4	5	6	1	2	3	4	5	6	1	2	3	4	5	6
OCTree	57.91	68.95	75.12	77.24	80.29	84.99	27.93	30.15	39.02	43.34	47.39	52.89	24.63	26.89	34.78	40.12	44.16	49.45
	±0.52	±0.63	±0.71	±0.68	±0.75	±0.25	±0.42	±0.45	±0.52	±0.56	±0.61	±0.65	±0.38	±0.41	±0.48	±0.53	±0.57	±0.62
CQFormer	49.52	52.53	62.91	71.34	75.49	79.89	19.92	20.01	29.95	32.81	39.62	41.89	19.91	20.01	29.95	32.81	39.65	41.89
	±0.48	±0.51	±0.58	±0.65	±0.69	±0.73	±0.35	±0.36	±0.45	±0.47	±0.52	±0.54	±0.34	±0.35	±0.44	±0.46	±0.51	±0.53

Appendix

Table 6: Performance comparison for different methods across varying color bits on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets. Mean and standard deviation are computed from 5 independent runs.

Color bit	1	2	3	4	5
No using	76.80	87.05	89.22	91.33	92.13
using	79.90	89.15	91.02	93.34	94.39

Appendix

Table 7: Effect of using minimizing the edge distribution differences on CIFAR-10.

	ShufffleNet, CIFAR-10					MobileNet-v2, CIFAR-10
	80%	83%	87.5%	92%	96%	80%	83%	87.5%	92%	96%
Random	83.05	80.29	72.18	70.34	60.18	81.25	78.09	70.38	68.04	57.88
EL2N	69.20	65.32	23.68	22.13	20.19	67.20	63.02	21.88	20.13	18.19
AUM	54.64	45.49	27.78	20.62	17.35	55.84	46.59	27.78	23.30	19.05
CCS AUM	82.14	80.32	77.51	65.94	63.45	80.64	78.02	75.01	63.44	61.15
TDDS	83.08	81.65	80.42	66.55	62.98	81.28	79.15	77.92	65.05	59.48

Appendix

Table 8: Comparison of dataset pruning algorithms and our algorithm on CIFAR-10 with ShufffleNet and MobileNet-v2.

	CIFAR-10					CIFAR-100
	80%	83%	87.5%	92%	96%	80%	83%	87.5%	92%	96%
Random	82.26	79.18	77.09	69.27	64.18	33.83	30.96	26.19	25.61	22.68
EL2N	59.13	55.15	27.77	24.97	21.64	22.49	14.55	11.67	9.18	8.76
AUM	65.61	56.51	28.65	23.22	20.17	22.58	15.32	12.88	10.15	6.08
CCS AUM	84.38	80.82	75.18	71.22	67.01	59.18	55.35	51.16	27.13	22.84
Ours	87.66	84.91	80.82	76.83	70.54	68.38	65.19	63.72	47.19	35.42

Appendix

Table 9: Comparison of dataset pruning algorithms and our algorithm on Transformer-Based Model. For CIFAR-10, we tested our algorithms on Swin Transformer; For CIFAR-100, we tested our algorithms on ViT-Small.

Color Bits	1	2	3	4	5
PCA	50.11	61.78	70.23	73.45	78.47
Gray Level Co-occurrence Matrix	70.15	83.17	85.49	86.17	90.75
Color Coherence Vector	74.15	86.97	87.19	89.77	92.35
RGB Color Histograms	75.95	86.18	88.49	90.17	91.45
Shallow-layer Feature Map	79.90	89.15	91.02	93.34	94.39

Appendix

Table 10: Different feature extraction strategies on CIFAR-10 trained on ResNet-18.

Color Bits	1	2	3	4	5
Mid-layer2 feature maps	69.15	79.81	81.09	83.76	87.75
Mid-layer1 feature maps	74.15	86.97	87.19	89.77	92.35
Final-layer Feature Map	42.10	53.78	66.39	75.15	80.44
Shallow-layer Feature Map	79.90	89.15	91.02	93.34	94.39

Appendix

Table 11: Comparison of Different layers on CIFAR-10 trained on ResNet-18.

Color Bits	1	2	3	4	5
80%	75.95	87.17	88.11	89.98	93.05
20%	77.31	87.79	88.23	90.55	91.47
70%	76.15	86.97	87.41	90.99	92.55
60%	76.85	87.17	88.89	91.97	93.01
30%	77.95	88.18	89.19	92.17	93.45
50%	79.90	89.15	91.02	93.34	94.39

Appendix

Table 12: Utilizing Grad-CAM++ attention maps to maintain the pixel distribution ratio

k_{Gra}\%

on CIFAR-10 and ResNet-18.

Table 13: Comparison of different dataset pruning algorithms on MS COCO.

Method	80%	87.5%	92%	96%
Random	16.8	39.4	50.2	119.8
CCS	15.7	37.8	52.8	121.1
EL2N	18.1	45.1	61.4	138.3
Ours	11.9	21.3	42.2	113.5

Appendix

Table 14: Comparison of dataset pruning algorithms on the CIFAR-10 PSLD generative task ( Pandey and Mandt, 2023 ) . Results are evaluated by FID (

\downarrow

), where lower values indicate better performance.

Table 15: Comparison of our DCQ with different dataset compression algorithms.

	CIFAR-10		CIFAR-100
	IPC=10	IPC=50	IPC=10	IPC=50
AutoPalette	74.3	79.4	52.6	53.3
Ours	77.9	82.3	58.9	61.7

Appendix

Table 16: Comparison between our algorithm and AutoPalette on CIFAR-10 and CIFAR-100 (ConvNetD3)

Method	80%	87.5%	92%	96%
Original setting	94.39	91.02	89.15	79.90
Original setting + ColorJitter	95.02	92.44	90.45	80.95
Original setting + GaussianBlur	95.47	92.95	91.47	81.75

Appendix

Table 17: Combining other dataset augmentation algorithms on CIFAR-10 (ResNet-18) under 1-bit quantization.

Method	80%	87.5%	92%	96%
Random	85.95	71.28	68.19	64.55
CCS	88.98	74.17	70.07	67.17
EL2N	67.32	23.98	16.31	14.85
Ours	91.16	76.22	72.34	70.17

Appendix

Table 18: Different dataset pruning algorithms on a fine-grained recognition task (Stanford Cars, API-Net ( Zhuang et al., 2020 ) ). Full dataset performance = 95.3%.

Method	80%	87.5%	92%	96%
Random	76.51	67.18	62.05	49.36
CCS	78.51	69.35	64.16	51.28
TDDS	77.44	67.99	63.66	50.85
Ours	82.26	73.44	68.34	59.17

Appendix

Table 19: Evaluation of dataset pruning algorithms on CIFAR-10 (ResNet-18) with 30% symmetric random label noise. Full dataset accuracy = 83.06%.

Number of Epochs	Number of Training Iterations	Acc (%)
50	9,750	75.01
100	19,500	76.19
125	24,375	77.69
150	29,250	78.57
200	39,000	79.90

Appendix

Table 20: Different number of training epochs on CIFAR-10 (ResNet-18) under 1-bit color quantization

Dataset	#Classes (C)	Images per Class (IPC)	Total Images	Complexity
ImageNet-1K	1,000	$\approx$ 1200	$\approx$ 1.2M	$O(1{,}000\times 1200\times\text{BackboneFLOPs})$
ImageNet-21K	21,000	$\approx$ 700	$\approx$ 14.7M	$O(21{,}000\times 700\times\text{BackboneFLOPs})$

Appendix

Table 21: Dataset statistics including class count, image distribution, total size, and complexity.

Dataset	Palette-Learning	Attention-Computation	Differentiable-Refinement	Total Time
ImageNet-1K	108 mins	16 mins	30 mins	154 mins
CIFAR-10/100	20 seconds	2 seconds	30 seconds	52 seconds

Appendix

Table 22: Time cost of generating different datasets with color quantization (sufficient memory).

Panel of Figure 7 — Figure 7: Visualization of different datasets under different color quantization ratios. Left column: CIFAR-10 images; Right column: CIFAR-100 images. From top to bottom: original images, 5-bits quantization (32 colors), and 1-bit quantization (2 colors).

Panel of Figure 8 — Figure 8: Visualization of Tiny-ImageNet under different color quantization ratios. From top to bottom: original images, 5-bits quantization (32 colors), and 1-bit quantization (2 colors).

Panel of Figure 9 — Figure 9: Visualization of different datasets under different color quantization ratios. Left column: CIFAR-10 images; Right column: CIFAR-100 images. From top to bottom: 2-bit quantization (4 colors), and 3-bit quantization (8 colors).

Panel of Figure 10 — Figure 10: Visualization of Tiny-ImageNet under different color quantization ratios. From top to bottom: 2-bit quantization (4 colors) and 3-bit quantization (8 colors).

Post Training Quantization for Efficient Dataset Condensation
Mar 7, 2026Linh-Tam Tran, Sung-Ho BaeDataset Condensation Post-Training Quantization
Kyung Hee University, Republic of Korea
From Fewer Samples to Fewer Bits: Reframing Dataset Distillation as Joint Optimization of Precision and Compactness
Mar 2, 2026My H. Dinh, Aditya Sant, Akshay Malhotra +2Quantization-Aware Distillation Dataset Distillation
InterDigital Communications, Inc.
Condensing Large-Scale Datasets Directly with Minimal Information Loss
Jul 1, 2026Xinyi Shang, Peng Sun, Bei Shi +2Distribution Shift Distribution Matching
University College London, United Kingdom · Zhejiang University, China · Westlake University, China +1

Mar 7, 2026J/K move · Enter open · S save

Post Training Quantization for Efficient Dataset Condensation

Linh-Tam Tran, Sung-Ho Bae

Kyung Hee University, Republic of Korea

Dataset Condensation Post-Training Quantization Low-Bit Quantization Dataset Distillation

Open paper AskPDF arXiv

Dataset Condensation (DC) distills knowledge from large datasets into smaller ones, accelerating training and reducing storage requirements. However, despite notable progress, prior methods have largely overlooked the potential of quantization for further reducing storage costs. In this paper, we take the first step to explore post-training quantization in dataset condensation, demonstrating its effectiveness in reducing storage size while maintaining representation quality without requiring expensive training cost. However, we find that at extremely low bit-widths (e.g., 2-bit), conventional quantization leads to substantial degradation in representation quality, negatively impacting the networks trained on these data. To address this, we propose a novel patch-based post-training quantization approach that ensures localized quantization with minimal loss of information. To reduce the overhead of quantization parameters, especially for small patch sizes, we employ quantization-aware clustering to identify similar patches and subsequently aggregate them for efficient quantization. Furthermore, we introduce a refinement module to align the distribution between original images and their dequantized counterparts, compensating for quantization errors. Our method is a plug-and-play framework that can be applied to synthetic images generated by various DC methods. Extensive experiments across diverse benchmarks including CIFAR-10/100, Tiny ImageNet, and ImageNet subsets demonstrate that our method consistently outperforms prior works under the same storage constraints. Notably, our method nearly doubles the test accuracy of existing methods at extreme compression regimes (e.g., 26.0%

\rightarrow

54.1% for DM at IPC=1), while operating directly on 2-bit images without additional distillation.

Abstract

Explore similar work

Post Training Quantization for Efficient Dataset Condensation

From Fewer Samples to Fewer Bits: Reframing Dataset Distillation as Joint Optimization of Precision and Compactness

Condensing Large-Scale Datasets Directly with Minimal Information Loss