Quantized Vision-Language Models for Damage Assessment: A Comparative Study of LLaVA-1.5-7B Quantization Levels

Takato Yasuno

cs.CVMar 24, 2026

Quantized Vision-Language Models for Damage Assessment: A Comparative Study of LLaVA-1.5-7B Quantization Levels

Authors: Takato Yasuno

Abstract

Bridge infrastructure inspection is a critical but labor-intensive task requiring expert assessment of structural damage such as rebar exposure, cracking, and corrosion. This paper presents a comprehensive study of quantized Vision-Language Models (VLMs) for automated bridge damage assessment, focusing on the trade-offs between description quality, inference speed, and resource requirements. We develop an end-to-end pipeline combining LLaVA-1.5-7B for visual damage analysis, structured JSON extraction, and rule-based priority scoring. To enable deployment on consumer-grade GPUs, we conduct a systematic comparison of three quantization levels: Q4_K_M, Q5_K_M, and Q8_0 across 254 rebar exposure images. We introduce a 5-point quality evaluation framework assessing damage type recognition, severity classification. Our results demonstrate that Q5_K_M achieves the optimal balance: quality score 3.18 $\pm$ 1.35/5.0, inference time 5.67s/image, and 0.56 quality/sec efficiency -- 8.5% higher quality than Q4_K_M with only 4.5% speed reduction, while matching Q8_0's quality with 25% faster inference. Statistical analysis reveals Q5_K_M exhibits the weakest text-quality correlation (-0.148), indicating consistent performance regardless of description length.

Figures & tables

Panel of Figure 1 — Figure 1: End-to-end pipeline architecture. This study focuses on Stage 2 (Vision Analysis) quantization comparison, while other stages remain constant.

Component	Points	Criteria
Damage Types	2.0	Recognition of: crack, rebar exposure, corrosion, spalling (0.5 pt each)
Severity Level	1.0	Classification: minor / moderate / severe
Location Info	1.0	Spatial description: top, bottom, left, right, beam, column, etc.
Extent Info	1.0	Coverage quantification: local, widespread, partial, percentage, dimensions
Total	5.0	Comprehensive assessment

Table 1: 5-Point Quality Evaluation Framework

Quantization	Size	Precision	Format
Q4_K_M	4.1 GB	4-bit mixed	GGUF
Q5_K_M	4.8 GB	5-bit mixed	GGUF
Q8_0	7.2 GB	8-bit uniform	GGUF
FP16 baseline	14.0 GB	16-bit float	safetensors

Table 2: Quantization Configurations

Figure 2: Comprehensive performance comparison across quantization levels (N=254). Top-left: Average inference time with error bars. Top-right: Model size vs total processing time. Bottom-left: Quality score distributions with baseline reference. Bottom-right: Summary metrics table.

$\star$ Recommended for production deployment (best quality-speed balance)
Quantization	Model Size	Init Time	Avg Inference	Quality Score	Text Length
Q4_K_M	4.1 GB	3.6 s	5.43 s	2.93 $\pm$ 1.53	168 $\pm$ 41 chars
Q5_K_M $\star$	4.8 GB	4.5 s	5.67 s	3.18 $\pm$ 1.35	160 $\pm$ 37 chars
Q8_0	7.2 GB	5.9 s	7.63 s	3.27 $\pm$ 1.39	162 $\pm$ 39 chars
Success rate: 254/254 (100%)

Table 3: Quantization Comparison Results (N=254)

Figure 3: Statistical analysis of quantization comparison (N=254). Top row: Quality score, text length, and inference time distributions (violin plots). Middle row: Box plots and scatter plots showing relationships. Bottom: Detailed metrics summary table.

∗∗ p $<$ 0.01; ∗ p $<$ 0.05
Comparison	U Statistic	p-value
Q5_K_M vs Q4_K_M	34,822.5	0.0591
Q8_0 vs Q5_K_M	33,870.0	0.1627
Q8_0 vs Q4_K_M	36,289.5	0.0069 ∗∗

Table 4: Pairwise Statistical Tests

Quantization	Correlation	Interpretation
Q4_K_M	$-0.559$	Moderate negative
Q5_K_M	$-0.148$	Weak negative
Q8_0	$-0.393$	Moderate negative

Table 5: Text-Quality Correlation Analysis

Component (Max)	Q4_K_M	Q5_K_M	Q8_0
Damage Types (2.0)	1.15	1.28	1.31
Severity (1.0)	0.65	0.72	0.74
Location (1.0)	0.58	0.63	0.65
Extent (1.0)	0.55	0.55	0.57
Total (5.0)	2.93	3.18	3.27

Table 6: Quality Component Breakdown (Average Scores)

Panel of Figure 4 — Figure 4: Preprocessed input images (N=12) after NLM denoising, resizing to 640×480, and CLAHE enhancement. Parentheses indicate damage level and damage type recognized by the model.

Photo	Type	Severity	Key Features
Photo 1	crack	medium	—
Photo 2	rebar_exposure	high	Cracks limited to specific areas; Repair needed
Photo 3	crack	high	—
Photo 4	rebar_exposure	high	—
Photo 5	crack	high	Cracks; Loads exceeding structural strength
Photo 6	crack	high	—

Table 7: Structured damage fields extracted by GPT-OSS-Swallow-8B_Q5_K_M.

Photo	Score (0-1)	Urgency (1-5)	Timeline
Photo 1	0.692	Level 3	Planned repair (1-2 years)
Photo 2	1.0	Level 5	Immediate repair (critical)
Photo 3	0.952	Level 5	Immediate repair (critical)
Photo 4	1.0	Level 5	Immediate repair (critical)
Photo 5	0.712	Level 4	Early repair (6 months)
Photo 6	0.692	Level 3	Planned repair (1-2 years)

Table 8: Priority scoring results for twelve representative bridge damage images.

VLM Quantization

Vision-language model quantization reduces model precision to lower memory and computation while preserving multimodal capabilities.

See all 52 papers

Fine-Tuning Vision-Language Models for Understanding Current Damage and Scoring Priority with Quality Guard Agent
May 24, 2026Takato YasunoVLM Adaptation Structural Health Monitoring
Llama-Mobile: Efficient 2.7-Bit Quantization of VLMs
Aug 21, 2026Luka Ribar, Jeevan Bhoot, Douglas OrrVLM Quantization Efficient VLM Inference
Graphcore Research · Arm
SalQ-VLM: Fine-Grained Saliency-Guided Quantization for Vision-Language Models
Aug 5, 2025Yufei Xue, Yushi Huang, Lunjie Zhu +2Vision-Language Models VLM Quantization
Institute of Artificial Intelligence (TeleAI), China Telecom · The Hong Kong University of Science and Technology

May 24, 2026J/K move · Enter open · S save

Fine-Tuning Vision-Language Models for Understanding Current Damage and Scoring Priority with Quality Guard Agent

Takato Yasuno

VLM Adaptation Structural Health Monitoring

Open paper AskPDF arXiv

Bridge inspection in Japan requires mandatory visual assessments every five years, yet qualitative damage ratings (levels a-e) assigned by different engineers exhibit significant inter-rater variability -- a critical barrier to consistent infrastructure management. The aging of skilled engineers further threatens inspection capacity. This paper presents a methodology for automating bridge damage understanding and repair priority scoring using fine-tuned Vision-Language Models (VLMs). We fine-tune LLaVA-1.5-7B with QLoRA on up to 4,000 paired bridge damage images and inspection text records, then evaluate on a fixed test set of 800 images. The model outputs natural language descriptions identifying structural members and damage patterns, from which a rule-based scoring engine calculates a five-level repair priority index. A progressive training study (1k/2k/3k/4k samples) reveals that 2k training samples achieve near-optimal validation loss in only 2.9 hours of training; beyond 2k, validation loss improves by no more than 0.2% per doubling of training samples, exhibiting clear diminishing returns. Furthermore, semantic similarity on the held-out test set peaks at 3k (0.6909) and degrades at 4k (0.6739), indicating that quality-curated mid-scale data outperforms larger but noisier corpora. Inference optimization combining torch.compile() and batch processing (batch_size=8) achieves 10.06 seconds per image -- a 70.2% reduction over the unoptimized baseline. Our approach contributes to data governance in bridge inspection, reduces inter-rater variability, and provides AI-assisted triage to augment expert engineers in inspection workflows. Furthermore, we introduce a two-stage Quality Guard using a fine-tuned Swallow-8B SLM to reject low-quality VLM outputs before priority scoring, preventing spurious scores from damaged or unrecognised images.

Abstract

VLM Quantization

Explore similar work

Fine-Tuning Vision-Language Models for Understanding Current Damage and Scoring Priority with Quality Guard Agent

Llama-Mobile: Efficient 2.7-Bit Quantization of VLMs

SalQ-VLM: Fine-Grained Saliency-Guided Quantization for Vision-Language Models