Research map / Video / Video-language understanding

ViDiC: Video Difference Captioning

NeurIPS 2026 · Video › Video-language understanding · approach: Video LLMs with fine-grained understanding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces the Video Difference Captioning (ViDiC) task with the ViDiC-1K benchmark (1,000 video pairs, 3,720 checklist items) scored by a dual-checklist LLM-as-judge protocol, evaluates 17 MLLMs, and fine-tunes Qwen2.5-VL-7B on a 60k-pair training set.

Problem

Existing MLLMs struggle to perceive and describe fine-grained similarities and differences between video pairs, including motion, camera work, and editing changes; image difference captioning does not cover temporal dynamics and existing video editing evaluations focus on edit fidelity rather than d

Approach

Extends image difference captioning to video pairs and evaluates descriptions with a dual-checklist of binary similarity and difference questions, where a blind text-only LLM judge (GPT-5-mini) checks answers against human-verified ground truth; similarity questions penalize only hallucinated differences.

Key results

• Fine-tuned ViDiC-Qwen (Qwen2.5-VL-7B) improves average score from 38.68 to 50.43 on ViDiC-1K (+11.75), while the human baseline reaches 94.46.
• Gemini-2.5-Pro is the best model at 69.33 average, while the best open-source model Qwen3-VL-32B reaches 63.90.
• GPT-5-mini as judge reaches 95.22% average concordance with human annotators, versus 89.37% for DeepSeek-V3 and 87.23% for Qwen3-32B.

Surprising findings

• LLaVA-v1.6-Vicuna-7B scores 0.58 on Difference while scoring 93.79 on Similarity, indicating it essentially never reports differences and produces repetitive outputs on dual-video input.
• Thinking mode improves both difference and similarity scores for several models, and Qwen-series models are nearly order-insensitive while InternVL3.5 and Keye-VL-1.5 are not.

Limitations

• Judge is a text-only LLM relying on the generated caption, and the judge's accuracy depends on its agreement with human annotators (95% on a sample).
• Benchmark videos are partly synthetic (Veo3, CV and rendering edits), which may not reflect all real edit distributions.
• Only one model (Qwen2.5-VL-7B) is fine-tuned; the training set labels are generated by Qwen3-VL-32B.

Experimental setup

Datasets: Ditto-1M LMArena Miradata Ego-Exo4D MultiCamVideo Vript VidDiffBench IF-VidCap VACE PKU-DyMVHumans ToCaDa DVSC

Benchmarks: ViDiC-1K LVBench MMBench-Video Video-MMMU Video-MME

Metrics: accuracy difference accuracy similarity accuracy concordance rate with human

Baselines: Gemini-2.5-Pro Gemini-3.0-Flash Gemini-2.5-Flash GPT-5 Qwen3-VL Qwen2.5-VL-Instruct InternVL-3.5 Keye-VL-1.5 Mimo-VL-SFT Kimi-VL-A3B GLM-4.1V InternVideo2.5

Builds on: Spot-the-Diff VidDiffBench LLM-as-a-judge POPE HallusionBench

Closest papers by meaning

• VEU-Bench: Towards Comprehensive Understanding of Video Editing (CVPR 2025 · similarity 0.87)
• ViTeX-Bench: Benchmarking High-Fidelity Video Scene Text Editing (NeurIPS 2026 · similarity 0.87)
• AuroraCap: Efficient, Performant Video Detailed Captioning and a New Benchmark (ICLR 2025 · similarity 0.87)
• EditBoard: Towards a Comprehensive Evaluation Benchmark for Text-Based Video Editing Models (AAAI 2025 · similarity 0.87)
• Vript: A Video Is Worth Thousands of Words (NeurIPS 2024 · similarity 0.86)
• VE-Bench: Subjective-Aligned Benchmark Suite for Text-Driven Video Editing Quality Assessment (AAAI 2025 · similarity 0.86)