Platform For AI:Deploy a LoRA SD model with Kohya_ss on EAS

Prerequisites

• Activate EAS and create a default workspace. For more information, see Activate PAI and create a default workspace.

• For RAM users deploying the model, grant permissions to manage EAS. For more information, see Cloud product dependencies and permissions: EAS.

• Create an OSS bucket and folders in the same region to store training materials, model files, and logs. For information about uploading files, see Upload files.

Prepare OSS storage structure

1. Log on to the OSS console and navigate to a bucket path in the same region. For example, oss://kohya-demo/kohya/.

2. In the current bucket path, create a project folder, for example, KaraDetroit_loar. In this folder, create three subfolders named Image, Log, and Model. Optionally upload a JSON configuration file to this folder.

   

   • Image: Source files for training materials.

   • Model: Model files after training completes.

   • Log: Training logs.

   • SS_config.json: JSON file for batch parameter configuration (optional). Modify parameters such as folder paths and output model names. For more information, see GitHub. For a sample file, see SS_config.json.

3. Upload prepared images to the Image folder. This topic uses a sample package containing images and description files: 100_pic.zip. Download and decompress the file, then upload the resulting folder to OSS. The result is shown in the following figure.

   

   Important

   • Supported image formats are .png, .jpg, .jpeg, .webp, and .bmp.

   • Each image must have a description file with the same name and a .txt file extension. The description must be on the first line. If an image has multiple descriptions, separate them with commas.

   • Image folder names must follow the format: number_name. For example, 100_pic. The name can be any string following OSS file naming rules. The number indicates repetitions for each image during training. This value is typically 100 or greater. Total training iterations generally exceed 1,500.

     • If the folder contains 10 images, each is trained 1500/10=150 times. Set the folder name number to "150".

     • If the folder contains 20 images, each is trained 1500/20=75 (<100) times. Set the folder name number to "100".

Deploy Kohya_ss service

1. Log on to the PAI console. Select a region on the top of the page. Then, select the desired workspace and click Elastic Algorithm Service (EAS).

2. Click Deploy Service. In the Custom Model Deployment section, click Custom Deployment.

3. On the Custom Deployment page, configure the parameters using the form or by specifying JSON configurations.

   Form configuration

   Parameter		Description
   Basic Information	Service Name	Custom service name. This topic uses kohya_ss_demo as an example.
   Environment Information	Deployment Method	Select Image-based Deployment and select Enable Web App.
   	Image Configuration	In the Alibaba Cloud Image list, select kohya_ss > kohya_ss:2.2. Note Versions iterate quickly. Select the latest image version during deployment.
   	Storage Mount	Select OSS as mount type and configure these parameters: Uri: Select an OSS path in the same region. This topic uses oss://kohya-demo/kohya/. Mount Path: Customize the path. This topic uses /workspace. Important Turn off the Read-Only switch. Otherwise, model files cannot be written to OSS.
   	Command to Run	System automatically configures the run command after selecting the runtime image. This topic uses python -u kohya_gui.py --listen=0.0.0.0 --server_port=8000 --headless. --listen: Binds program to the specified local IP address to accept external requests. --server_port: Listening port number.
   Resource Information	Resource Type	Select Public Resources.
   	Deployment	Resource specifications: ml.gu7i.c16m60.1-gu30 GPU is recommended for cost-effectiveness. This topic uses the current minimum specification, ml.gu7i.c8m30.1-gu30.

   JSON configuration

   In the Service Configurations section, click Edit and enter the configuration in JSON format.

   Sample JSON configuration:

   Important

   In the following example, modify the configurations on line 4 ("name") and line 18 ("oss") based on your requirements.

   {
       "metadata":
       {
           "name": "kohya_ss_demo",
           "instance": 1,
           "enable_webservice": true
       },
       "cloud":
       {
           "computing":
           {
               "instance_type": "ecs.gn6e-c12g1.12xlarge",
               "instances": null
           }
       },
       "storage": [
       {
           "oss":
           {
               "path": "oss://kohya-demo/kohya/",
               "readOnly": false
           },
           "properties":
           {
               "resource_type": "model"
           },
           "mount_path": "/workspace"
       }],
       "containers": [
       {
           "image": "eas-registry-vpc.cn-hangzhou.cr.aliyuncs.com/pai-eas/kohya_ss:1.2",
           "script": "python -u kohya_gui.py --listen=0.0.0.0 --server_port=8000 --headless",
           "port": 8000
       }]
   }

4. Review the configuration and click Deploy. Model deployment takes a few minutes. The service is deployed when Service Status changes to Running.

Train LoRA model

1. On the Overview page of the service, click Web applications to open the Kohya_ss service page.

2. Select LoRA (LoRA).

   

3. Set Configuration file. This step is optional. Skip if no SS_config.json file is available.

   

   Note

   Configuration file path consists of Mount Path from the Form configuration section, the OSS folder path, and SS_config.json. In this topic, the path is /workspace/KaraDetroit_loar/SS_config.json.

4. Set SourceModel. This topic uses safetensors, which offers better security than checkpoint.

   

5. Set Folders parameters. Enter paths for Image, Log, and Model folders created in OSS and specify the output file name.

   

   Parameter	Description
   Image folder	Folder path containing training images. Set to Mount Path from the Form configuration section, concatenated with the Image path created in OSS. In this topic, the path is /workspace/KaraDetroit_loar/Image.
   Logging folder	Folder path for log output. Set to Mount Path from the Form configuration section, concatenated with the Log path created in OSS. In this topic, the path is /workspace/KaraDetroit_loar/Log.
   Output folder	Folder path for model output. Set to Mount Path from the Form configuration section, concatenated with the Model path created in OSS. In this topic, the path is /workspace/KaraDetroit_loar/Model.
   Model output name	Model output name. For example, my_model.

6. Set Parameters. For example values, see the SS_config.json file content in the Prepare OSS storage structure step.

   Parameter	Description
   LoRA Type	LoRA type. Options: LoCON: Adjusts every layer of SD, such as Res, Block, and Transformer. LoHA: Processes more information with the same size.
   LoRA network weights	LoRA network weights. For continuing training, select the last trained LoRA (optional).
   Train batch size	Training batch size. Larger values require more VRAM.
   Epoch	Number of training epochs. One epoch is one complete pass through all training data. Calculate this value manually. Generally: Total training iterations in Kohya = Number of training images × Repetitions × Number of epochs / Training batch size. Total training iterations in WebUI = Number of training images × Repetitions. When using class images, total training iterations double in both Kohya and WebUI. Number of model saves in Kohya is halved.
   Save every N epochs	Saves model every N epochs. For example, setting to 2 saves training result every 2 epochs.
   Caption Extension	File extension for caption files, for example, .txt (optional).
   Mixed precision	Mixed precision. Depends on graphics card performance. Options: no, fp16, and bf16. For graphics cards with more than 30 GB VRAM, set to bf16.
   Save precision	Save precision. Same as above.
   Number of CPU threads per core	Number of CPU threads per core. Mainly affects VRAM. Adjust based on purchased instance and needs.
   Learning rate	Learning rate. Default: 0.0001.
   LR Scheduler	Learning rate scheduler. Select a function such as cosine or cosine with restart as needed.
   LR Warmup (% of steps)	Number of prefetch steps. Adjust as needed. Default: 10. Select 0 if prefetch is not required.
   Optimizer	Optimizer. Select as needed. Default: AdamW8bit. DAdaptation represents automatic operation.
   Max Resolution	Maximum resolution. Set based on image properties.
   Network Rank (Dimension)	Model complexity. A value of 128 is sufficient for most scenarios.
   Network Alpha	Set to a value smaller than or equal to Network Rank (Dimension). Common setting: Network Rank 128 with Network Alpha 64.
   Convolution Rank (Dimension) & Convolution Alpha	Convolution degree. Defines the scope of LoRA's fine-tuning on the model. Adjust based on LoRA Type. Kohya official recommendations: LoCon: dim <= 64, alpha = 1 (or lower). LoHA: dim <= 32, alpha = 1.
   clip skip	Number of times the CLIP model is used. Range: 1 to 12. Smaller values produce images closer to the original input. For realistic models, select 1. For anime-style models, select 2.
   Sample every n epoch	Sample every N epochs. Saves a sample at specified epoch intervals.
   Sample Prompts	Sample prompts. Use a command with parameters: --n: Prompt, negative prompt. --w: Image width. --h: Image height. --d: Image seed. --l: Prompt relevance (cfg). --s: Iteration steps (steps).

7. At the bottom of the page, click Start training.

8. In the service list on the EAS page, click the service name to go to the service details page. Click Log to view training progress.

   

   When model saved appears in the log, training is complete.

   

9. After training completes, retrieve the trained LoRA model file from the configured model folder (Model), for example, my_model.safetensors.

   

Use trained LoRA model for AI art generation

After training a LoRA model, upload it to the corresponding SD WebUI directory and mount it to generate images. For more information about deploying Stable Diffusion service, see Deploy a Stable Diffusion service with EAS in 5 minutes to enable text-to-image generation.

The following sections describe uploading a LoRA model file to Stable Diffusion WebUI.