LVM-Aesthetic Filter (DLC)

更新时间:
复制 MD 格式

LVM-Aesthetic Filter (DLC) removes low-quality videos from your training dataset by scoring each video's visual aesthetics. Only MP4 videos are supported. Use this component to clean video training data before training video generation models.

How it works

The component samples a fixed number of frames from each video at equal intervals, scores each frame using an aesthetic model, and computes the mean score across all sampled frames. Videos whose mean aesthetic score falls outside the configured minimum–maximum range are filtered out.

You must select a GPU instance type for the algorithm.

Supported computing resources

Deep Learning Containers (DLC)

Inputs and outputs

Input ports

  • Read File Data: reads the Object Storage Service (OSS) path where the training data is stored.

  • Any component from LVM Data Processing (DLC) can also serve as the input.

For the input OSS path, configure the OSS Data Path parameter to specify either:

  • An OSS directory containing the video files, or

  • The video metadata file video_meta.jsonl generated by a previous run of this component.

Output port

The output port produces the filtering results, written to the OSS directory specified by Output File OSS Path.

Configure the component

Configure the component parameters in Machine Learning Designer.

Field settings

ParameterRequiredDefaultDescription
Video Data OSS PathNoThe OSS directory containing the video data. On the first run without an upstream component, specify this directory manually. The component writes a video_meta.jsonl metadata file to the parent directory of the path you specify. On subsequent runs, point this parameter to the video_meta.jsonl file directly instead of the raw directory.
Output File OSS PathYesThe OSS directory where filtering results are stored. The component writes three files: {name}.jsonl (filtered output), {name}_stats.jsonl (run statistics), and dj_run_yaml.yaml (the parameter configuration used during the run).
Output FilenameYesresult.jsonlThe base file name for the output file.

Parameter settings

ParameterRequiredDefaultDescription
Minimum Aesthetic ScoreYes0.3Videos with a mean aesthetic score below this threshold are filtered out.
Maximum Aesthetic ScoreYes1Videos with a mean aesthetic score above this threshold are filtered out.
Number of Sampled FramesYes3The number of frames sampled per video. Frames are drawn at equal intervals across the video duration.

Execution tuning

ParameterRequiredDefaultDescription
Number of ProcessesYes4The number of parallel worker processes.

Select resource group

OptionRequiredDescription
Public Resource GroupNoSelect the instance type (CPU or GPU) and virtual private cloud (VPC). You must select the GPU instance type for the algorithm.
Dedicated resource groupNoSpecify the number of vCPUs, memory, shared memory, and GPUs to allocate.
Maximum Running Duration (seconds)NoThe maximum time the component is allowed to run. The job is terminated if this limit is exceeded.

Usage notes

Reuse the metadata file

After the first run, a video_meta.jsonl file is written to the parent directory of Video Data OSS Path. On subsequent runs against the same dataset, point Video Data OSS Path to this file directly instead of the raw directory.