LVM-Watermark Filter (DLC)

更新时间:
复制 MD 格式

LVM-Watermark Filter (DLC) is a Platform for AI (PAI) component that filters watermarked videos from training datasets. It runs as a Deep Learning Containers (DLC) job and processes only MP4 videos.

How it works

The component samples frames from each video at uniform intervals, runs a watermark detection model on each sampled frame to produce a probability score, and then compares that score against the configured maximum watermark probability threshold. Videos that exceed the threshold are filtered out. The results are written to the output path for use in downstream training pipelines.

The typical use case is cleaning datasets before training video generation models.

Important

The watermark detection model requires a GPU instance. Select a GPU instance type when configuring the resource group.

Note

Only MP4 videos are supported.

Supported computing resources

Deep Learning Containers (DLC)

Inputs and outputs

Input

Connect one of the following as the upstream component:

  • Read File Data — reads the Object Storage Service (OSS) path where training data is stored. Configure the OSS Data Path parameter to point to the OSS directory containing the video files, or to a previously generated video_meta.jsonl metadata file.

  • Any other component from LVM Data Processing (DLC).

Output

Filtering results written to the OSS directory specified by Output File OSS Path. The output directory contains the following files:

FileDescription
{name}.jsonlOutput file. The filename is controlled by the Output Filename parameter.
{name}_stats.jsonlState file.
dj_run_yaml.yamlParameter configuration used when the algorithm ran.

Configure the component

Configure the following parameters in Machine Learning Designer.

Field settings

ParameterRequiredDefaultDescription
Video Data OSS PathNoThe OSS directory containing the video files. If no upstream component is connected on the first run, set this parameter manually. Running the component generates video_meta.jsonl in the parent directory of the path you specify. On subsequent runs, point this parameter directly to video_meta.jsonl to skip re-scanning.
Output File OSS PathYesThe OSS directory where filtering results are written.
Output FilenameYesresult.jsonlThe base filename for the output and statistics files.

Parameter settings

ParameterRequiredDefaultDescription
Watermark ProbabilityYes0.8The maximum watermark probability. Videos whose watermark probability exceeds this threshold are filtered out.
Number of Sampled FramesYes1The number of frames sampled from each video for analysis. Frames are selected at uniform intervals across the video duration.

Execution tuning

ParameterRequiredDefaultDescription
Number of ProcessesYes4The number of parallel processes used during filtering.

Resource group

ParameterRequiredDefaultDescription
Public Resource GroupNoThe instance type (CPU or GPU) and virtual private cloud (VPC) to use. Select a GPU instance type — the watermark detection algorithm requires a GPU.
Dedicated resource groupNoThe number of vCPUs, memory, shared memory, and GPUs to allocate.
Maximum Running Duration (seconds)NoThe maximum time the component can run. If this limit is exceeded, the job is terminated.