FIFA: Unified Faithfulness Evaluation Framework for Text-to-Video and Video-to-Text Generation
This addresses the issue of unreliable content generation in video AI systems, which is incremental as it builds on existing evaluation methods by extending them to multiple tasks and open-ended responses.
The paper tackles the problem of hallucinations in Video Multimodal Large Language Models (VideoMLLMs) for text-to-video and video-to-text generation by proposing FIFA, a unified faithfulness evaluation framework that extracts facts, models dependencies, and verifies them, showing it aligns more closely with human judgment and improves factual consistency.
Video Multimodal Large Language Models (VideoMLLMs) have achieved remarkable progress in both Video-to-Text and Text-to-Video tasks. However, they often suffer fro hallucinations, generating content that contradicts the visual input. Existing evaluation methods are limited to one task (e.g., V2T) and also fail to assess hallucinations in open-ended, free-form responses. To address this gap, we propose FIFA, a unified FaIthFulness evAluation framework that extracts comprehensive descriptive facts, models their semantic dependencies via a Spatio-Temporal Semantic Dependency Graph, and verifies them using VideoQA models. We further introduce Post-Correction, a tool-based correction framework that revises hallucinated content. Extensive experiments demonstrate that FIFA aligns more closely with human judgment than existing evaluation methods, and that Post-Correction effectively improves factual consistency in both text and video generation.