UGC-VideoCaptioner: An Omni UGC Video Detail Caption Model and New Benchmarks

TL;DR

The paper addresses the need for true omnimodal video understanding in real-world UGC by introducing the UGC-VideoCap benchmark and the UGC-VideoCaptioner-3B model. It pairs a richly annotated, audio-visual short-video dataset with a two-stage distillation pipeline (teacher auto-labeling via Gemini-2.5-Flash followed by SFT and GRPO-based RLHF) to achieve data-efficient, high-quality omnimodal captions. Empirical results show balanced audio-visual captioning capabilities and demonstrate the effectiveness of the SFT+RL approach over plain data scaling, establishing a practical framework for future omnimodal captioning in unconstrained video content. The work provides both a valuable benchmark and a scalable training paradigm that can catalyze progress in real-world multimodal video understanding across UGC and cinema contexts.

Abstract

Real-world user-generated videos, especially on platforms like TikTok, often feature rich and intertwined audio visual content. However, existing video captioning benchmarks and models remain predominantly visual centric, overlooking the crucial role of audio in conveying scene dynamics, speaker intent, and narrative context. This lack of omni datasets and lightweight, capable models hampers progress in fine grained, multimodal video understanding. To address these challenges, we introduce UGC-VideoCap, a new benchmark and model framework specifically designed for detailed omnimodal captioning of short form user-generated videos. Unlike prior datasets, UGC-VideoCap emphasizes balanced integration of audio and visual modalities, featuring 1000 TikTok videos annotated through a structured three stage human-in-the-loop pipeline covering audio only, visual only, and joint audio visual semantics. The benchmark also includes 4000 carefully crafted QA pairs probing both unimodal and cross modal understanding. Alongside the dataset, we propose UGC-VideoCaptioner(3B), a 3B parameter captioning model distilled from Gemini 2.5 Flash. Using a novel two-stage training strategy supervised fine tuning followed by Group Relative Policy Optimization (GRPO), our approach enables efficient adaptation from limited data while maintaining competitive performance. Together, our benchmark and model offer a high-quality foundation and a data-efficient solution for advancing omnimodal video captioning in unconstrained real-world UGC settings.

Figures (4)

• Figure 1: Benchmark Collation Pipeline. This pipeline unifies disparate raw video annotation data into a standardised format for consistent processing. QA pairs are then generated from question templates with fixed rules. To ensure quality, manual validation is performed at all key stages, with multiple people cycling through low quality and ambiguous annotated content to assess whether to re-labelling.
• Figure 2: Demonstration of tasks for UGC-VideoCap benchmark. It has many meta information for audio and visual. And we select some of them to be our benchmark QA pairs which are visual detail description, audio track detail description and final detail caption.
• Figure 3: Evaluation on UGC-VideoCap Benchmark. We provided a detailed evaluation of the audio and visual, and finally designed a detailed caption with audio and video. Regarding the OCR score we are considering the average of the 4 NLP metrics (BLEU-1, ROUGE-1, ROUGE-2, and ROUGE-L), scaled to percentage for fair comparison in the average calculation.
• Figure 4: Left:Distillation Stage. We use Gemini-2.5-flash to generate 20k omni video detail caption. Right:Reinforcement Learning Stage. We use GRPO algorithm and new caption reward model to enhance the detail caption ability.