LVM-Image-Watermark Filter (DLC)-Platform For AI(PAI)-阿里云帮助中心

Watermarked images in training data cause generative models to reproduce watermark artifacts in their outputs. The LVM-Image-Watermark Filter (DLC) component scores each image for watermark probability and removes images that exceed the threshold you set, keeping watermark artifacts out of your image generation model training pipelines.

Supported computing resources

Deep Learning Containers (DLC)

How it works

For each image, the component calculates a watermark probability score. Images with a score above the Watermark Probability threshold are filtered out. The remaining images and their metadata are written to the output path you specify.

Typical pipeline position:

Read File Data (or any image preprocessing component)
    → LVM-Image-Watermark Filter (DLC)
        → Image generation model training component

Inputs and outputs

Input ports

The component accepts input from any of the following:

Read File Data component — reads the Object Storage Service (OSS) path where training data is stored
Image Data OSS Path parameter — specify an OSS directory or an image metadata file directly (see the parameter table below)
Any image data preprocessing component

Output port

Filtering results are written to the OSS path specified by OSS Directory for Saving OutputData. The output includes three files:

File	Description
`{name}.jsonl`	The filtered dataset. The filename is set by Output Filename. Pass this file to the downstream training component.
`{name}_stats.jsonl`	The state file.
`dj_run_yaml.yaml`	The parameter configuration file used when the algorithm runs.

Configure the component

Configure the LVM-Image-Watermark Filter (DLC) component in Machine Learning Designer. The following table describes all parameters.

Tab	Parameter	Required	Description	Default
Field Settings	Image Data OSS Path	No	The OSS directory where image data is stored, or an image metadata file. If no upstream component is connected on the first run, select the OSS directory manually. The component generates `meta.jsonl` in the parent directory of the path you specify. On subsequent runs, select `meta.jsonl` directly instead of the directory.	—
Field Settings	OSS Directory for Saving OutputData	No	The OSS directory where filtering results are stored. See the output file descriptions above.	—
Field Settings	Output Filename	No	The filename for the filtered dataset (the `{name}.jsonl` output file).	`result.jsonl`
Parameter Settings	Watermark Probability	Yes	The maximum watermark probability. Images with a score above this value are filtered out. Lower the threshold to remove more images; raise it to keep more.	`0.8`
Execution Tuning	Number of Processes	No	The number of processes.	`4`
	Single-node/Multi-node	No	Single-node or multi-node mode. Default value: single-node. Multi-node mode supports only the Ray framework.	single-node
	Select Resource Group	No	Select a public resource group or a resource quota (if distributed mode is used, you must set resource parameters for the head and worker nodes separately): Public resource group: You must select a GPU instance type for the algorithm. Resource quota: select the number of GPUs, number of vCPUs, memory (GB), and shared memory (GB).	—
	Maximum Running Duration (seconds)	No	The maximum time the component can run. The job is terminated if this limit is exceeded.	—

Important

You must select a GPU instance type when configuring the resource group.