AI Video Quality Assessment Classification Tool

Uses a large model to score AI-generated videos across six dimensions, identify major issues such as taking off in place, clipping, and subject jumps, and automatically classify and archive them based on thresholds.

Bug report

Loading tool, please wait...

If the online tool fails to load or run, try the desktop tool.https://tools.yikeaigc.com/

Tool Usage

Return to old version

A desktop version is available with more powerful, stable features. Open desktop version

Instructions

Software Usage Instructions

Configure API access: In "API Settings", select the native Gemini format, enter the Base URL and API Key for NewAPI or a compatible platform, and choose a model that supports video understanding (such as gemini-3-pro-preview).
Add videos to evaluate: Click "Add File" or "Add Folder" to select the MP4 videos to review. Recursive subdirectory scanning is supported, and you can also drag files directly into the window.
Set evaluation parameters:
- Score threshold: Set an overall score threshold between 0 and 100. Videos with scores not lower than the threshold are considered qualified.
- Number of concurrent threads: Adjust based on API rate limits and network conditions. Commonly 2-5 threads.
- Request timeout: The maximum waiting time in seconds for a single evaluation. For large files, it is recommended to increase this appropriately.
Select output directory: Specify the root directory for saving. The tool will automatically create two subdirectories under it: "Qualified" and "Unqualified", while preserving the original relative path structure.
Configure additional options:
- Generate HTML report: After processing, summarize the scores and issues of all videos into a visual report.
- Resume evaluation: Match existing CSV records based on file fingerprints and configuration, and skip videos that have already been evaluated.
- Preserve original structure: In recursive mode, archive files according to their relative paths from the source.
Start evaluation: Click the "Start Evaluation" button. The tool will send the video to the large model via inlineData and parse the returned six-dimensional scores and issue list.
View and export results: After processing is complete, you can view the score and critical issue description for each video in the real-time log, and obtain CSV details, individual JSON reports, and an HTML summary in the output directory.

FAQ

A: The tool scores six dimensions from 0 to 10: motion naturalness, physical plausibility, common-sense plausibility, temporal consistency, narrative logic, and visual quality. It also provides an overall quality score from 0 to 100 and a one-sentence summary to comprehensively reflect the quality of AI-generated videos.

A: Commonly covered critical issues include stiff or jerky motion, actions that violate physical laws such as taking off from a standstill, behaviors that violate common sense such as vehicles moving backward, subject or background jumps, clipping, abnormal limbs, blurry images, noise, compression artifacts, incoherent narrative logic, and more. Each issue includes a timestamp and a text description.

A: The current version supports videos in MP4 format. Because the video is embedded into the API request using inlineData, it is recommended to keep each video under 15MB. Files larger than this should be compressed or trimmed before being submitted for evaluation.

A: You can set an overall score threshold between 0 and 100. Videos with an overall score not lower than the threshold will be copied or moved to the "Passed" directory, while those below the threshold will go to the "Failed" directory. If preserving the original structure is enabled, the classified directories will retain the video's path hierarchy relative to the source directory, making it easier to locate files later.

A: The tool has built-in breakpoint resume for evaluation. Completed videos are written to the CSV with a unique key generated from parameters such as the file fingerprint, model, and threshold. When evaluation is started again, if the parameters are the same, the source file has not been modified, and the target file still exists, it will be skipped automatically and will not consume API quota again.

A: This usually does not mean the video itself scored zero, but that the API did not actually pass the video to the model. Please confirm that the Base URL uses the v1beta path or the generateContent endpoint, that the selected model supports video understanding, and that this capability has been enabled on the NewAPI/MagicAPI platform. Also check whether the file size exceeds the inline limit, and compress the video before retrying if necessary.

A: Each video generates a JSON report with the same name, containing the overall score, six-dimensional scores, issue list, and summary. The results of all videos are summarized into a CSV table, including fields such as status, source file, target file, scores, number of issues, and time taken. If HTML reports are enabled, a styled summary web page will also be generated in the output directory for quick browsing.