TLDR: A new fine-tuning method called IOVQA (Integer-only VQA) significantly improves Vision Language Models (VLMs) for video quality assessment. It converts human-assigned decimal quality scores into integers (10-50 range) and uses an “integer-only mask” during training, focusing the model on precise numerical predictions. This approach leverages VLMs’ strengths in handling discrete values, leading to more accurate and human-aligned video quality evaluations, demonstrated by a 3rd place ranking in a major challenge.
In the rapidly evolving world of artificial intelligence, Vision Language Models (VLMs) are becoming increasingly vital for evaluating visual content, especially in applications like assessing video quality and consistency. However, existing methods often struggle with accuracy and efficient calculations, which can limit how well these models focus on key evaluation indicators.
To tackle these challenges, researchers have introduced a novel fine-tuning approach called IOVQA, which stands for Integer-only VQA. This method is specifically designed to boost the performance of VLMs in video quality assessment tasks. The core innovation of IOVQA lies in how it constructs labels for training and its unique loss calculation mechanism.
During the creation of the dataset, IOVQA restricts the model’s output to integers within a specific range, from 10 to 50. This ensures numerical stability and involves converting original decimal-based Mean Opinion Scores (MOS) – which are human-assigned quality ratings – into integers before they are used as labels. This conversion is crucial because large language models, which are at the heart of VLMs, inherently perform better when predicting integer values rather than decimals. Decimal numbers require additional tokens (like the decimal point), which can extend the prediction chain and amplify errors.
Furthermore, IOVQA introduces a ‘target-mask strategy’. When the model calculates its errors (loss), only the first two-digit integer of the label is considered. This clever approach forces the model to concentrate on learning the most critical numerical components of the evaluation, reducing noise from less relevant parts of the output.
The researchers fine-tuned the Qwen2.5-VL model using this specially constructed dataset. Experimental results have shown that IOVQA significantly improves the model’s accuracy and consistency in video quality assessment. In fact, it achieved an impressive 3rd place ranking in the VQualA 2025 GenAI-Bench AIGC Video Quality Assessment Challenge – Track I. This highlights the effectiveness of simply using integer labels during fine-tuning, offering a powerful new idea for optimizing VLMs in scenarios that require quantitative evaluation.
This approach stands in contrast to some traditional metrics that might only provide a single overall score without capturing the multi-dimensional nature of human assessment, or methods that generate free-text explanations which can lead to inconsistencies. By constraining the model to generate only numerical scores, IOVQA ensures consistency and better alignment with human perception.
The Qwen2.5-VL model was chosen as the base for its advanced capabilities, including enhanced visual recognition, precise object positioning, and its ability to produce structured outputs. These features are particularly beneficial for tasks like video quality assessment, where fine-grained spatial and temporal understanding is key.
Also Read:
- Enhancing Visual Perception with Decoupled Learning: Introducing DeCLIP
- PMTFR: A Novel Framework for Enhanced Composed Image Retrieval
The success of IOVQA demonstrates that by aligning the label format and loss computation with the natural capabilities of VLMs, it’s possible to achieve more accurate and robust video quality assessments that closely mirror human judgment. For more details, you can refer to the original research paper here.
