Reference Fidelity in AI-Generated Short-Form Video Across Four Prompt Packages

Abstract

Generative video systems can produce short clips from textual and visual instructions, yet their ability to preserve the content of a human reference remains uncertain. This exploratory study examines 160 videos generated with Doubao-Seedance-2.0 from 40 human-made short-form videos drawn from Douyin and Xiaohongshu. The references were divided evenly among humor, informational, demonstration, and storytelling content. Qwen3.8-Max analyzed each reference and produced four prompt packages labeled vague, vivid, keyword, and image-based. The fourth package also contained timestamped reference frames, frame descriptions, identity constraints, and sequence instructions, so it is treated as a multimodal reference condition rather than a pure image condition. Each prompt was used once under the same default settings, and no output was edited after generation. Metadata inspection covered the complete corpus. All 40 groups received a standardized frame review, and eight groups were examined at three temporal positions. Outputs usually retained the source topic and central action. Fidelity was stronger for one-subject situations than for multi-step procedures, factual labels, comic timing, or longer causal sequences. The multimodal package appeared closer to source composition in several cases, yet the one-shot design and unequal prompt information prevent a causal ranking. The study therefore evaluates reference fidelity rather than virality or prompt efficacy.