AVE-Compass

---

Can audio-visual editing models correctly follow cross-modal instructions?

AVE-Compass overview

Dataset Statistics

AVE-Compass dataset statistics

Project page Paper citation Dataset GitHub code

**Work overview.** AVE-Compass is a diagnostic benchmark for free-form audio-video editing, built from 145 curated source videos, 196 audio-visually coupled instructions, and 2,688 fine-grained checklist items. Its evaluation separates Instruction Following, Fidelity Preserving, Realism, and Editing Intent, combining MLLM-as-Judge evaluation with automated cross-modal, video, and audio metrics to expose incomplete edits, non-target drift, perceptual artifacts, and audio-visual misalignment. **Repository overview.** This repository provides the AVE-Compass evaluation pipeline, configuration files, input templates, metric implementations, and AVE-Agent. The evaluator processes source videos, edit instructions, checklists, and model-generated edited videos with MLLM and objective metrics, while the [AVE-Compass dataset](https://huggingface.co/datasets/NJU-LINK/AVE-Compass) provides the benchmark samples. AVE-Agent provides a planning and self-reflection baseline for complex editing instructions.

Evaluation Code · Key Findings · AVE-Agent

--- ## What We Evaluate AVE-Compass contains 145 curated source videos, 196 human-verified editing instructions, 2,688 checklist items, and 28 editing operation types across joint, speech, video-only, and audio-only editing. The benchmark reports four complementary dimensions: | Dimension | Description | | --- | --- | | **Instruction Following** | Whether the requested edit is correctly executed | | **Fidelity Preserving** | Whether non-target visual and audio content remains faithful to the source | | **Editing Intent** | Whether the output both follows the instruction and preserves non-target content | | **Realism** | Whether the edited audio-video result is natural and coherent | The evaluation package also reports automated cross-modal, video, and audio metrics, together with source-preservation diagnostics for single-modality edits.

AVE-Compass evaluation matrix

## Leaderboard Models are ranked by **Overall Editing Intent**, the primary metric of AVE-Compass. Scores are reported on a 0-100 scale, and higher is better. | Rank | Model | Overall | Video | Audio | | ---: | --- | ---: | ---: | ---: | | 1 | **AVE-Agent (Wan)** | **59.8** | **66.7** | **50.2** | | 2 | Wan2.7 | 42.4 | 60.1 | 24.8 | | 3 | HappyHorse | 41.3 | 56.7 | 18.8 | | 4 | Gemini-Omni* | 38.0 | 56.1 | 10.0 | | 5 | Seedance | 26.6 | 36.1 | 13.5 | | 6 | LTX2 | 15.2 | 10.7 | 26.4 | *Gemini-Omni misses 16 speech edits due to content moderation. ## Evaluation Code ### 1. Configure Environment Install the dependencies and model weights required by the selected metrics, and make sure `ffmpeg` and `ffprobe` are available on `PATH`. MLLM evaluation uses the official Gemini API: ```bash export GEMINI_API_KEY=your_key export DEFAULT_MODEL=your_model ``` ```yaml modules: {objective: true, llm: true} llm: {checklist: true, realism: true} ``` ### 2. Prepare Inputs Download the released source videos, instructions, and checklists into the three default directories: ```bash hf download NJU-LINK/AVE-Compass --type dataset \ --include "videos/*" \ --include "edit_instructions/*" \ --include "checklists/*" \ --local-dir input ``` Place the videos produced by the model under evaluation in `edited_videos/`: ```text input/ |- videos/ # downloaded |- edit_instructions/ # downloaded |- checklists/ # downloaded `- edited_videos/.mp4 # provided by the user ``` ### 3. Run ```text bash run_eval.sh [input_root] [model_name] [output_path] [config_path] [device] ``` For example: ```bash bash run_eval.sh \ input \ my_model \ output \ configs/path.yml \ cuda ``` The five positional parameters default to `input`, `model`, `output`, `configs/path.yml`, and `cuda`, respectively. Use `cpu` as the last parameter for CPU-compatible metrics. ### 4. Outputs ```text output/ |- _eval__eval_results.json |- _eval__eval_results.csv `- / `- / |- eval_result.json |- objective/ `- llm/ ``` The CSV contains one row per evaluated instruction; the JSON additionally retains per-question judgments, metric details, warnings, and aggregate summaries. ## Key Findings - Current editors often struggle to satisfy instruction execution and non-target preservation at the same time. - A model that leaves the input unchanged can receive deceptively strong preservation scores, so edit response must be reported separately. - Audio-side execution and temporal audio-visual alignment remain common failure modes. - AVE-Agent improves Editing Intent and Instruction Following, with particularly strong gains in audio-side execution and audio-visual synchronization. ## AVE-Agent This repository also includes **AVE-Agent**, a modular audio-video editing agent that plans dependent subtasks, dispatches modality-specific tools, and refines failed edits through evaluator feedback. See the [`agent/` directory](agent/) for setup, architecture, and usage. ## Repository Layout ```text ./ |- README.md |- input/ | |- videos/ | |- edit_instructions/ | |- checklists/ | `- edited_videos/ |- run_eval.sh |- configs/ |- metrics/ |- modules/ |- models/ |- tests/ |- agent/ `- docs/assets/ ``` ## Reproducibility Notes - Input samples are paired by their shared ``, while the source video is resolved from the instruction JSON. - Missing or disabled objective metrics remain unset and are not converted into fabricated zero scores. - MLLM evaluation uses the official Gemini API. Run the evaluation checks with: ```bash python -m pytest tests ``` The tests cover input pairing, metric routing, missing-value handling, checklist aggregation, and Realism score normalization. ## Citation ```bibtex @article{wen2026avecompass, title = {AVE-Compass: Towards Holistic Evaluation for Audio-Video Editing Abilities}, author = {Wen, Yuqing and Huang, Yukai and Xie, Qianqian and Wu, Jiangtao and Lin, Yibin and Gu, Yikai and Chen, Jialu and Zhang, Yuanxing and Liu, Jiaheng}, year = {2026} } ```