Platform For AI:PAI ArtLab Kohya

Prerequisites

Before you begin, make sure you have:

• Enabled PAI ArtLab and granted the required permissions. See Enable PAI ArtLab and grant permissions

• (Optional) Claimed a free trial or coupon, or purchased a resource plan. See PAI ArtLab billing. Check the validity period to use your resources before they expire

How it works

Training a LoRA model with Kohya (Exclusive Edition) follows four stages:

1. Create a dataset — Upload training images into a named folder.

2. Caption images — Use the WD14 model to auto-generate text descriptions for each image.

3. Train the model — Configure and launch a LoRA training job. Monitor the loss value to assess training quality.

4. Evaluate the output — Run the trained LoRA in Stable Diffusion (Shared Edition) and compare results with an X/Y/Z plot.

The example in this guide trains an oil painting style model using 15 landscape images at 768 × 768 pixels.

Step 1: Create a dataset

1. Log in to PAI ArtLab. In the upper-right corner, hover over the icon and select China (Shanghai).

2. On the Dataset page, click Create Dataset and enter a name.

3. Open the dataset, click Create Folder, and enter a folder name. Folder names must follow the format Number_CustomName, where the number controls how many times images in the folder repeat during training. For example, 30_test repeats each image 30 times.

4. Upload your training images to the folder. Image quality requirements:

   • Use more than 15 clear images.

   • For LoRA training on the sd1.5 base model, 512 × 512 or 512 × 768 pixels is sufficient — higher resolutions are unnecessary.

   • Avoid images with watermarks, low definition, unusual lighting, complex or unrecognizable content, or unusual angles.

Step 2: Caption images

Captioning generates a text description for each image. The WD14 model reads each image and creates a prompt describing its content. You can review and edit captions afterward if needed.

1. On the Toolbox page, click the Kohya (Exclusive Edition) card to open the tool.

2. Go to the > Captioning tab and configure the following parameters.

   Parameter	Description
   Image folder to caption	Select the folder you created. If it doesn't appear in the drop-down list, enter the path manually — for example, /data-oss/datasets/test/30_test.
   Undesired Tags	Enter any tags you want to exclude from the generated captions.
   Prefix to add to WD14 caption	Enter the LoRA trigger word. Use the format DatasetName + Number — for example, test1.

3. Click Caption images. Captioning takes 2–3 minutes. When captioning done appears in the log, captioning is complete.

4. On the Datasets page, open your folder and click any image to view its caption. Edit the caption text if needed.

Step 3: Train the model

Select a base model

On the Model > Model Scope page, select a Checkpoint base model and add it to My Models.

Method	When to use	Steps
Preset model (recommended)	You want a platform-provided model, such as sd1.5 xl	Select the model directly from the Model Scope page.
Custom model	You have your own Checkpoint model	Upload a base model or add an existing model to My Models first.

For a custom model, set Model Quick Pick to custom. In the Pretrained model name or path field, enter /data-oss/models/Stable-diffusion, append /, and then select the Checkpoint model you added or uploaded to My Models.

Step 4: Evaluate the model

Use an X/Y/Z plot to compare different training checkpoints and LoRA strength values side by side.

1. On the Model > My Models page, click the icon on a model card to add both the Checkpoint model and the trained LoRA model to Stable Diffusion (Shared Edition).

2. On the Toolbox page, click the Stable Diffusion (Shared Edition) card.

3. Click the icon next to Stable Diffusion Model and select your Checkpoint model.

4. On the Text-to-image tab, go to the Generation tab and configure:

   Parameter	Value
   Steps	30
   Script	X/Y/Z plot
   X Type	Prompt S/R
   X Values	NUM,000001,000002,000003
   Y Type	Prompt S/R
   Y Values	STRENGTH,0.3,0.5,0.6,0.7,0.8,0.9,1

5. On the LoRA tab, click Refresh and select the LoRA model you trained. If your LoRA model isn't listed, select any trained LoRA model and update the prompt to reference your model. For example, change <lora:test-000002:1> to <lora:test-NUM:STRENGTH>.

6. Enter the prompts:

   Field	Value
   Positive Prompt	test1, outdoors, sky, day, cloud, water, tree, blue sky, no humans, traditional media, grass, building, nature, scenery, house, castle,
   Negative Prompt	lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit,fewer digits, cropped, worst quality, low quality,normal quality, jpeg artifacts, signature,watermark, username, blurry,(worst quality:1.4),(low quality:1.4), (monochrome:1.1), Eagetive,

7. Click Generate.

The X/Y/Z plot shows results for all combinations of checkpoint epochs (X axis) and LoRA strength values (Y axis). Use it to identify the checkpoint and strength that best match your target style.

What's next

• To learn about billing and resource plans, see PAI ArtLab billing.