Immersive Computer Graphics (CGs) rendering has become ubiquitous in modern daily life. However, comprehensively evaluating CG quality remains challenging for two reasons: First, existing CG datasets lack systematic descriptions of rendering quality; and second existing CG quality assessment methods cannot provide reasonable text-based explanations. To address these issues, we first identify six key perceptual dimensions of CG quality from the user perspective and construct a dataset of 3500 CG images with corresponding quality descriptions. Each description covers CG style, content, and perceived quality along the selected dimensions. Furthermore, we use a subset of the dataset to build several question-answer benchmarks based on the descriptions in order to evaluate the responses of existing Vision Language Models (VLMs). We find that current VLMs are not sufficiently accurate in judging fine-grained CG quality, but that descriptions of visually similar images can significantly improve a VLM's understanding of a given CG image. Motivated by this observation, we adopt retrieval-augmented generation and propose a two-stream retrieval framework that effectively enhances the CG quality assessment capabilities of VLMs. Experiments on several representative VLMs demonstrate that our method substantially improves their performance on CG quality assessment.
Figures & tables
Figure 1: An example illustrating the capability of the proposed R4-CGQA. R4-CGQA not only makes the responses of large models more concise and accurate, but also effectively unleashes their potential on the CGQA task.
Figure 2: We prepare 10 CG images with associated questions and answers (Q1: multiple-choice questions and Q2: yes-or-no questions) and evaluate answer accuracy on LLaVA-13B [ 18 ] . Dark bars indicate accuracy when the evaluation is directly provided by the VLM; light bars indicate accuracy when reference descriptions from visually similar CG images are provided to the VLM.
Figure 3: CG Image Quality Perception Factors.
Figure 4: Overview of dataset characteristics. (a) Example image–description pairs illustrating diverse visual content. (b) Distribution of description lengths across the dataset. (c) Word cloud visualization of the most frequent descriptive terms, highlighting key perceptual attributes such as texture, light, and realism. (d) Top 30 most frequent words and their corresponding frequencies.
Figure 5: GPT-4o is used to generate three types of questions for the validation and testing sets.
Figure 6: Overview of R4-CGQA. A query CG image and a CG library are embedded in a content space, and a local neighborhood around the query is formed. Within this neighborhood, quality embeddings are computed and combined to select the most similar example, whose textual description is used as additional context when querying the VLM assistant together with the input CG, producing a final quality judgement and explanation. The CG library is built by our base set and it is scalable; ASS denotes average similarity search.
VLMs
Original
R4-CGQA
↑ Average
Choice
Yes-or-no
Q&A
Choice
Yes-or-no
Q&A
LLava-1.6-7B
51.70%
49.73%
2.20
58.06% (+6.36%)
58.84% (+9.11%)
2.50 (+6.00%)
7.16%
LLava-1.6-13B
53.96%
51.34%
2.22
61.43% (+7.47%)
60.37% (+9.03%)
2.56 (+6.80%)
7.77%
Llama 3.2-V-11B
64.59%
56.87%
1.93
67.28% (+2.69%)
67.26% (+10.39%)
2.31 (+7.60%)
6.89%
Gemma3-4B
66.03%
53.55%
1.05
66.82% (+0.79%)
65.22% (+11.67%)
2.32 (+25.40%)
12.62%
MiniCPM-V-8B
60.05%
53.47%
1.92
67.63% (+7.58%)
61.98% (+8.51%)
2.34 (+8.40%)
8.16%
Table 1: Evaluating state-of-the-art VLMs on our testing set, the “Original” part denotes the results of directly using VLMs to answer questions, and the R4-CGQA denotes the results of our approach. Scores in red denote the improvement.
Figure 7: Qualitative example of the R4-CGQA system. The left panel shows the query CG; the right panel shows the top- 5 retrieved images with their CLIP similarity ( sim1 ), REIQA-based similarity ( sim2 ) and averaged score ( avg ). The bottom text illustrates VLM answers to two questions, combining a scalar quality judgment with a detailed natural language explanation.
Llava-1.6-7B
Llama3.2-V-11B
Config.
Choice
Yes-or-no
Q&A
Choice
Yes-or-no
Q&A
Base
50.1%
48.8%
2.23
65.3%
55.8%
1.94
w/o. quality
56.8%
59.0%
2.31
61.0%
68.3%
2.21
w/o. content
57.0%
60.4%
2.29
65.2%
68.9%
2.30
Full
59.8%
59.9%
2.38
66.7%
69.0%
2.43
Table 2: Ablation study of content and quality based retrieval. “w/o.” denotes without.
Question
Base
Multi-image (only)
Multi-image + R4-CGQA
R4-CGQA
Choice
70.8%
68.5%
71.0%
73.5%
Yes-or-no
68.3%
71.6%
71.1%
71.7%
Table 3: Explore multi-image performance on Pixtral-12B; T=0.9 and K=7 .
Question
K=1
K=3
K=5
K=7
K=9
Choice
57.1%
59.0%
59.8%
58.7%
57.2%
Yes-or-no
57.5%
58.0%
59.9%
56.5%
57.0%
Table 4: Explore the impact of different K values on the performance of Llava-1.6-7B responses; T is set to 0.8 .
Figure 8: The influence of threshold T and the accuracy rate of responses was tested on LlavaNext-8B, and we investigate K=5 and K=7 .
School of Engineering, University of Galway, Ireland · Lero the Research Ireland Centre for Software, Ireland · State Key Laboratory of Extreme Photonics and Instrumentation, Zhejiang University, Hangzhou, China.