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

Prerequisites

• Activate EAS and create a default workspace. Activate PAI and create a default workspace.

• If a RAM user deploys the model, grant EAS management permissions. Cloud product dependencies and permissions: EAS.

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

Set up OSS storage

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

2. Create a project folder (example: KaraDetroit_loar) with three subfolders: 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). Configure folder paths, output model names, and other parameters. GitHub. Sample: SS_config.json.

3. Upload prepared images to the Image folder. This example uses a sample package: 100_pic.zip. Download, decompress, and upload the folder to OSS.

   

   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 (example: 100_pic). The number sets repetitions per image during training (typically 100+). Total training iterations should 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 name. Example: kohya_ss_demo.
   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 Select the latest version during deployment.
   	Storage Mount	Select OSS as mount type and configure these parameters: Uri: OSS path in the same region. Example: oss://kohya-demo/kohya/. Mount Path: Custom path. Example: /workspace. Important Turn off the Read-Only switch. Otherwise, model files cannot be written to OSS.
   	Command to Run	Auto-configured after image selection. Example: python -u kohya_gui.py --listen=0.0.0.0 --server_port=8000 --headless. --listen: Binds to the specified IP address for external requests. --server_port: Listening port number.
   Resource Information	Resource Type	Select Public Resources.
   	Deployment	Recommended: ml.gu7i.c16m60.1-gu30 for cost-effectiveness. This example uses 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. Deployment takes a few minutes. The service is ready 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. (Optional) Set Configuration file. Skip if no SS_config.json file is available.

   

   Note

   The path combines Mount Path from Form configuration, the OSS folder path, and the file name. Example: /workspace/KaraDetroit_loar/SS_config.json.

4. Set SourceModel. safetensors is recommended over checkpoint for better security.

   

5. Set Folders. Enter the Image, Log, and Model paths from OSS, and specify the output file name.

   

   Parameter	Description
   Image folder	Training images folder. Combine Mount Path from Form configuration with the OSS Image path. Example: /workspace/KaraDetroit_loar/Image.
   Logging folder	Log output folder. Combine Mount Path from Form configuration with the OSS Log path. Example: /workspace/KaraDetroit_loar/Log.
   Output folder	Model output folder. Combine Mount Path from Form configuration with the OSS Model path. Example: /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 weights. To resume training, select the previous LoRA output (optional).
   Train batch size	Training batch size. Larger values require more VRAM.
   Epoch	Number of epochs (one full pass through all training data). Calculate manually: 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. With class images, total iterations double in both Kohya and WebUI, and model saves in Kohya are halved.
   Save every N epochs	Saves the model every N epochs. Example: set to 2 to save after every 2 epochs.
   Caption Extension	File extension for caption files, for example, .txt (optional).
   Mixed precision	Mixed precision. Options: no, fp16, bf16. Use bf16 for GPUs with 30+ GB VRAM.
   Save precision	Save precision. Same as above.
   Number of CPU threads per core	CPU threads per core. Affects VRAM usage. Adjust based on your instance specs.
   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)	Warmup steps. Default: 10. Set to 0 to skip warmup.
   Optimizer	Optimizer. Default: AdamW8bit. DAdaptation enables automatic tuning.
   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	Must not exceed Network Rank (Dimension). Common setting: Rank 128, Alpha 64.
   Convolution Rank (Dimension) & Convolution Alpha	Convolution degree. Controls the fine-tuning scope. Adjust based on LoRA Type. Kohya official recommendations: LoCon: dim <= 64, alpha = 1 (or lower). LoHA: dim <= 32, alpha = 1.
   clip skip	CLIP skip count. Range: 1 to 12. Lower values produce images closer to the input prompt. For realistic models, select 1. For anime-style models, select 2.
   Sample every n epoch	Generates a sample image every N epochs.
   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. Click the service name in the EAS service list to open the details page, then click Log to view training progress.

   

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

   

9. Retrieve the trained LoRA model from the Model folder. Example: my_model.safetensors.

   

Generate images with the trained LoRA model

Upload the trained LoRA model to the SD WebUI LoRA directory for image generation. Deploy a Stable Diffusion service with EAS in 5 minutes to enable text-to-image generation.

Upload the LoRA model file to the SD WebUI LoRA directory using one of the following methods.