Thinking with Frames: Generative Video Distortion Evaluation via Frame Reward Model

TL;DR

REACT introduces a frame-level reward model designed to evaluate structural distortions in generative videos, addressing a gap where existing reward models emphasize aesthetics or temporal coherence over correctness of object structure. The approach builds a taxonomy-guided, large-scale frame dataset, an efficient Chain-of-Thought data synthesis pipeline, and a two-stage training regime combining supervised fine-tuning with masked learning and reinforcement learning via Group Relative Policy Optimization to produce both point-wise distortion scores and attribution labels. A dynamic, two-stage frame sampling strategy during inference further concentrates evaluation on likely distorted frames, and REACT-Bench provides a dedicated benchmark for distortion-focused assessment. Empirical results show that REACT outperforms state-of-the-art video evaluators in human-preference alignment and distortion recognition, with ablations underscoring the importance of the taxonomy, CoT synthesis, and GRPO-based fine-tuning. The work lays groundwork for more reliable feedback in T2V systems and points to future extensions into spatio-temporal reasoning to detect temporal distortions that require temporal context.

Abstract

Recent advances in video reward models and post-training strategies have improved text-to-video (T2V) generation. While these models typically assess visual quality, motion quality, and text alignment, they often overlook key structural distortions, such as abnormal object appearances and interactions, which can degrade the overall quality of the generative video. To address this gap, we introduce REACT, a frame-level reward model designed specifically for structural distortions evaluation in generative videos. REACT assigns point-wise scores and attribution labels by reasoning over video frames, focusing on recognizing distortions. To support this, we construct a large-scale human preference dataset, annotated based on our proposed taxonomy of structural distortions, and generate additional data using a efficient Chain-of-Thought (CoT) synthesis pipeline. REACT is trained with a two-stage framework: ((1) supervised fine-tuning with masked loss for domain knowledge injection, followed by (2) reinforcement learning with Group Relative Policy Optimization (GRPO) and pairwise rewards to enhance reasoning capability and align output scores with human preferences. During inference, a dynamic sampling mechanism is introduced to focus on frames most likely to exhibit distortion. We also present REACT-Bench, a benchmark for generative video distortion evaluation. Experimental results demonstrate that REACT complements existing reward models in assessing structutal distortion, achieving both accurate quantitative evaluations and interpretable attribution analysis.

Figures (7)

• Figure 1: Comparison of REACT with SOTA Video and Image Evaluators. (a) While existing evaluators tend to assign high scores based on aesthetics and temporal consistency, even in the presence of structural defects, our REACT model outperforms them by accurately identifying structural distortions in generative videos and providing more reliable scores (b) While image evaluators excel in recognizing image artifacts, they struggle to detect distortions in generative video frames. In contrast, REACT demonstrates superior performance in recognizing and evaluating structural distortions in video frames.
• Figure 2: Overview of REACT: Frame-Level Reward Model for Structural Distortion Evaluation. (a) We first construct a large-scale annotated dataset, including human preference and attribution labels, based on our proposed detailed taxonomy of structural distortions. Additionally, we generate Chain-of-Thought (CoT) data using our proposed efficient CoT Synthesis pipeline. (b) We then train REACT based on Qwen2.5-VL-7B using a two-stage training framework. In the supervised fine-tuning (SFT) stage, a masked loss is applied to improve domain knowledge injection. In the reinforcement learning (RL) stage, pair-wise rewards are introduced to align the output point-wise scores of REACT with human preferences. (3) Finally, frames most likely to exhibit distortions are adaptively selected with a dynamic sampling mechanism, enabling flexible analysis within fixed frame sampling constraints.
• Figure 3: Detailed Explanation of Our Proposed Taxonomy of Structural Distortions in Generative Videos. Representative examples for each distortion category are also provided.
• Figure 4: Text Prompt for Our REACT in Human Preference Alignment Task.
• Figure 5: Text Prompt for Our REACT in Distortion Recognition Task.