PixelPrune: Pixel-Level Adaptive Visual Token Reduction via Predictive Coding

Nan Wang; Zhiwei Jin; Chen Chen; Haonan Lu

cs.CVApr 1, 2026

PixelPrune: Pixel-Level Adaptive Visual Token Reduction via Predictive Coding

Authors: Nan Wang, Zhiwei Jin, Chen Chen, Haonan Lu

Organizations: OPPO AI Center

Abstract

Document understanding and GUI interaction are among the highest-value applications of Vision-Language Models (VLMs), yet they impose exceptionally heavy computational burden: fine-grained text and small UI elements demand high-resolution inputs that produce tens of thousands of visual tokens. We observe that this cost is largely wasteful -- across document and GUI benchmarks, only 22--71% of image patches are pixel-unique, the rest being exact duplicates of another patch in the same image. We propose PixelPrune, which exploits this pixel-level redundancy through predictive-coding-based compression, pruning redundant patches before the Vision Transformer (ViT) encoder. Because it operates in pixel space prior to any neural computation, PixelPrune accelerates both the ViT encoder and the downstream LLM, covering the full inference pipeline. The method is training-free, requires no learnable parameters, and supports pixel-lossless compression ( $τ{=}0$ ) as well as controlled lossy compression ( $τ{>}0$ ). Experiments across three model scales and document and GUI benchmarks show that PixelPrune maintains competitive task accuracy while delivering up to 4.2 $\times$ inference speedup and 1.9 $\times$ training acceleration. Code is available at https://github.com/OPPO-Mente-Lab/PixelPrune.

Figures & tables

Figure 1: PixelPrune patch selection on a document image ( left ) and GUI screenshot ( right ). Kept patches are shown in original color; dropped patches are grayed out. Token reduction: 70% (document) and 93% (GUI).

Figure 2: Prefill latency breakdown across Qwen3-VL scales (2B, 4B, 8B) at five resolutions (256 2 –4096 2 , log scale). Each bar splits into Vision Encoder (red, including ViT and Patch Merger) and LLM Prefill (blue). At 4096 2 , the vision encoder accounts for 86%, 72%, and 75% of total prefill time for the 2B, 4B, and 8B models respectively.

Figure 3: Illustration of PixelPrune’s Pred-2D prediction. For each target patch

X

, three causal neighbors—

A

(left),

B

(upper),

C

(upper-left)—determine the predicted patch

\hat{X}

. The rule selects the neighbor most likely to match

X

: when the upper and upper-left patches agree, the target is more likely to follow the left neighbor, and vice versa. If

X

matches its prediction, it is omitted; otherwise, it is retained.

Model	Method	OCRBench	DocVQA	InfoVQA	ChartQA	AI2D	MML-Doc	olmOCR	Avg
		(66.1%)	(70.1%)	(76.7%)	(73.1%)	(69.2%)	(50.3%)	(56.9%)
Qwen3-VL-2B	Full	86.9	92.7	72.3	78.8	76.6	28.1	48.9	69.2
	Resize	86.1	92.3	72.0	76.9	76.5	26.3	47.8	68.3
	Random	76.0	66.8	62.9	57.3	72.2	17.0	34.7	55.3
	ConnComp	77.2	80.5	66.1	62.8	73.9	18.4	37.2	59.4
	PixelPrune	87.3	92.0	71.5	77.0	75.8	26.5	48.0	68.3

Table 1: Training-free results on Document Understanding Benchmarks (%). Numbers in parentheses show PixelPrune’s dataset-level retain ratio; all baselines use the same per-image token budget.

	ScreenSpot V2				ScreenSpot Pro
Method	Web	Mobile	Desktop	Avg	Dev.	Creative	CAD	Sci.	Office	OS	Avg
	(39.7%)	(49.1%)	(61.9%)		(64.3%)	(59.0%)	(65.6%)	(26.3%)	(56.4%)	(61.9%)
Full	88.1	92.0	91.6	90.6	45.8	41.6	40.2	53.2	70.4	50.0	50.2
Resize	87.4	91.6	88.0	89.0	44.1	38.4	37.2	37.4	62.2	41.3	43.4
Random	48.5	62.7	68.0	59.7	29.4	24.6	26.4	12.6	37.8	29.6	26.8
ConnComp	69.8	82.0	75.1	75.7	35.8	27.9	28.7	22.1	47.8	31.1	32.2

Table 2: Post-training results on GUI Understanding Benchmark (Qwen3-VL-2B). We report grounding accuracy (%). Numbers in parentheses show the dataset-level retain ratio. We additionally report PixelPrune (+KD) with knowledge distillation.

		ScreenSpot V2				ScreenSpot Pro
Train	Infer	Web	Mobile	Desktop	Avg	Dev.	Creative	CAD	Sci.	Office	OS	Avg
Full	Full	79.9	81.6	83.8	81.8	31.8	26.7	15.3	35.0	49.6	30.6	31.5
PixelPrune	PixelPrune	79.9	85.6	88.9	84.8	30.8	27.3	15.7	34.6	52.6	27.6	31.4
PixelPrune	Full	80.6	86.0	88.6	85.1	30.4	27.9	16.1	35.0	50.4	29.6	31.6

Table 3: Training from scratch results on GUI Understanding (Qwen3-VL-2B). “PixelPrune

\to

Full” = trained with PixelPrune, evaluated with full tokens.

		Selector	Retain	FLOPs	ViT Lat.	TTFT	KV Cache	FLOPs	TTFT
Dataset	Method	(ms)	(%)	(TFLOPs)	(ms)	(ms)	(GB)	Speedup	Speedup
MML-Doc	Full	–	100.0	2694	4898	18477	13.2	1.0 $\times$	1.0 $\times$
MML-Doc	PixelPrune	19.1	50.3	762	2143	6114	7.2	3.5 $\times$	3.0 $\times$
SS-Pro Sci.	Full	–	100.0	70	625	833	0.8	1.0 $\times$	1.0 $\times$
SS-Pro Sci.	PixelPrune	1.2	26.2	11	115	198	0.3	6.6 $\times$	4.2 $\times$

Table 4: Inference efficiency on Qwen3-VL-2B (batch size 1, H20 GPU). MML-Doc: MMLongBench-Doc; SS-Pro Sci.: ScreenSpot Pro Scientific. Selector: per-sample overhead of the pixel-level selector. Speedup factors are relative to the Full baseline.

Method	Retain	Forward (s)	Backward (s)	Peak Mem (GB)	Wall Time (h)
Full	100.0%	7.9	35.3	57.8	49.1
PixelPrune	40.8%	3.9	18.0	38.4	25.3
Speedup	–	2.0 $\times$	2.0 $\times$	$\downarrow$ 33.6%	1.9 $\times$

Table 5: Training efficiency measured during the from-scratch experiment in Section 4.4 . Speedup is relative to Full.

	Document						GUI
	DocVQA		InfoVQA		ChartQA		Web		Mobile		Desktop
Method	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.
Raster	70.1	92.0	77.8	71.4	76.7	77.6	42.5	87.2	50.5	91.8	64.4	90.1
Serpentine	70.0	92.1	77.4	71.2	74.9	77.1	41.8	88.6	50.1	91.2	64.1	88.9
Pred-2D	70.1	92.0	76.7	71.5	73.1	77.0	39.7	86.5	49.1	92.2	61.9	90.1

Table 6: Comparison of prediction strategies.

	Doc.			GUI
Depth	Doc	Info	Chart	Web	Mob.	Desk.
After ViT	92.2	71.0	76.8	88.1	92.2	90.4
After 17th	91.9	71.1	76.1	86.7	92.4	92.2
After 4th	92.2	71.4	77.0	89.9	92.6	89.8
After 0th	92.1	71.3	76.9	86.7	93.0	90.4
Before ViT	92.0	71.5	77.0	86.5	92.2	90.1

Table 7: Token reduction at different ViT depths.

		Document						GUI
		DocVQA		InfoVQA		ChartQA		Web		Mobile		Desktop
Distance	$\tau$	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.
Full (100%)		100	92.7	100	72.3	100	78.8	100	88.1	100	92.0	100	91.6
Exact ( $\tau$ =0)		70.1	92.0	76.7	71.5	73.1	77.0	39.7	87.6	49.1	92.4	61.9	91.6
MAE	0.005	61.8	92.0	66.7	70.4	53.1	74.4	36.0	87.4	35.9	92.2	42.7	91.6
	0.020	37.0	90.9	56.9	68.4	35.0	69.4	31.3	86.7	27.1	91.2	30.7	88.9

Table 8: Distance metrics and thresholds. Document = training-free; GUI = +KD model from Section 4.3 .

\tau

is normalized to

[0,1]

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Document Understanding
Benchmark	#Images	Mean Resolution	Duplicate Ratio (%)
OCRBench	1,000	615 ${\times}$ 732	38.5
DocVQA	1,286	1775 ${\times}$ 2082	33.4
InfoVQA	500	1141 ${\times}$ 2929	28.6
ChartQA	2,500	768 ${\times}$ 584	40.8
AI2D	1,204	609 ${\times}$ 486	37.7

Appendix

Table 9: Pixel-level patch redundancy and resolution statistics (patch size

32{\times}32

, exact match). Duplicate Ratio = fraction of non-unique patches (global dedup within each image).

Model	Method	OCRBench	DocVQA	InfoVQA	ChartQA	AI2D	MML-Doc	olmOCR	Avg
		(66.1%)	(70.1%)	(76.7%)	(73.1%)	(69.2%)	(50.3%)	(56.9%)
Qwen3.5-2B	Full	87.3	93.5	75.4	78.8	79.6	32.4	49.9	71.0
Qwen3.5-2B	PixelPrune	87.9	92.5	74.3	77.9	78.7	31.6	48.7	70.2
Qwen3.5-4B	Full	86.6	95.4	81.6	85.1	84.9	45.4	53.5	76.1
Qwen3.5-4B	PixelPrune	86.6	95.3	80.7	84.7	84.7	45.4	53.6	75.9
Qwen3.5-9B	Full	88.3	95.7	84.4	86.2	86.7	53.1	54.1	78.4

Appendix

Table 10: Training-free results on Document Understanding Benchmarks using Qwen3.5. Format follows Table 1 . Numbers in parentheses show the dataset-level retain ratio of PixelPrune.

		ScreenSpot V2				ScreenSpot Pro
Mode	Step	Web	Mobile	Desktop	Avg	Dev.	Cre.	CAD	Sci.	Office	OS	Avg
Full	400	66.1	78.2	75.1	73.2	14.7	14.1	5.0	20.5	22.6	10.7	14.6
	800	74.4	84.4	82.3	80.4	22.4	19.6	12.6	24.4	29.1	15.3	20.6
	1200	78.0	87.2	85.9	83.7	24.7	22.6	14.9	29.5	40.9	24.5	26.2
	1600	78.7	88.4	85.0	84.1	29.8	28.2	13.4	31.5	42.6	25.5	28.5
	2000	78.7	84.4	83.5	82.2	25.1	23.5	14.2	27.6	40.9	25.0	26.0

Appendix

Table 11: Training from scratch convergence: intermediate evaluation results every 400 steps. We report grounding accuracy (%) on ScreenSpot V2 (SSv2) and ScreenSpot Pro (SSPro). “Full” = trained and evaluated with all tokens; “PixelPrune” = trained and evaluated with compressed tokens.

	MMBench		MMMU		MMStar		RealWorldQA		HR-Bench		V*Bench
$\tau$	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.	Ret.	Acc.
Full (no dedup)	100	77.1	100	50.3	100	55.1	100	66.3	100	70.6	100	71.2
0 (exact)	94.7	77.1	80.1	49.2	81.2	55.1	97.7	66.3	96.3	71.0	99.7	71.2
0.005	94.2	76.9	79.2	49.7	80.5	55.1	97.6	66.1	94.9	70.8	99.3	71.2
0.02	91.8	76.9	76.1	49.1	77.2	55.3	96.6	65.8	84.6	71.4	94.0	73.8
0.05	88.0	76.5	72.4	50.3	73.3	55.0	89.2	66.1	77.3	72.0	87.7	72.8

Appendix

Table 12: General-domain evaluation on Qwen3-VL-2B (training-free, Max distance metric). Each benchmark reports retain ratio (%) and accuracy (%). Bold indicates the best accuracy among PixelPrune variants per benchmark.

Efficient VLM Inference

Efficient vision-language model inference reduces the compute, memory, latency, and energy needed to process images, video, and text.

See all 485 papers

VPRune: Efficient Training-free Pre-LLM Visual Token Pruning
Sep 21, 2026Guangchuan Lv, Dianxing Shi, Dingjie FuEfficient VLM Inference Large Vision-Language Models
Northeastern University · Beihang University · Huazhong University of Science and Technology
From Patching to Pruning Visual Computation in Vision Language Models
Oct 2, 2026Rahul Chowdhury, Timothy A Rupprecht, Xuan Shen +3Efficient VLM Inference Visual Token Pruning
Northeastern University, USA · EmbodyX Inc., USA · Zhejiang University, China +1
Language-Guided Token Compression with Reinforcement Learning in Large Vision-Language Models
Mar 11, 2026Sihan Cao, Jianwei Zhang, Pengcheng Zheng +7Efficient VLM Inference Visual Token Pruning
School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China · School of Robotics and Advanced Manufacture, Harbin Institute of Technology, Shenzhen, China · College of Information and Control Engineering, Xi’an University of Architecture and Technology, Xi’an, China

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

VPRune: Efficient Training-free Pre-LLM Visual Token Pruning

Guangchuan Lv, Dianxing Shi, Dingjie Fu

Northeastern University · Beihang University · Huazhong University of Science and Technology

Efficient VLM Inference Large Vision-Language Models Visual Token Pruning Visual Token Compression LVLM Pruning

Open paper AskPDF arXiv

Visual token pruning is a promising approach to reducing the inference cost of large vision-language models (LVLMs), yet aggressive token reduction often causes substantial performance degradation. We identify three key factors behind this degradation: text-guided selection bias, information loss from discarded tokens, and positional distortion caused by sequence compaction. Based on these observations, we propose VPRune, a training-free pre-LLM pruning framework consisting of visual-only diversity selection, similarity-guided token recycling, and position-preserving restoration. Experiments on FastVLM-1.5B across multiple vision-language benchmarks demonstrate that VPRune achieves a favorable accuracy--compression trade-off, with particularly pronounced advantages under aggressive compression. Furthermore, evaluations on edge-device show that VPRune effectively reduces end-to-end inference latency while maintaining superior task performance, demonstrating its practicality for resource-constrained LVLM deployment.

Abstract

Efficient VLM Inference

Explore similar work

VPRune: Efficient Training-free Pre-LLM Visual Token Pruning

From Patching to Pruning Visual Computation in Vision Language Models

Language-Guided Token Compression with Reinforcement Learning in Large Vision-Language Models