Skip to main content

Stable Diffusion Finetuning and Deployment


Recommended tutorials before starting to work with stable diffusion finetuning and deployment jobs.

1. Create and upload datasets

Firstly we need to upload your images to our platform before finetuning stable diffusion model using these images. Navigate through the Stochastic Platform in the Datasets side tab and select Add dataset. Next, add your dataset name, dataset description, choose Image as type and browse your images folder. Then click Upload to create an images dataset.

Images dataset upload

The number of images should be higher than 5 and lower than 30 and the original resolution should be higher than 512x512.

2. Create finetuning job

Navigate through the Jobs side tab and select Create job. You can select Stable Diffusion 1 or Stable Diffusion 2 as the initialized pretrained model for your finetuning job. Stable Diffusion 2 is chosen in this example.

Jobs side

Next, choose your uploaded images dataset and put your Class name. Class name of the images dataset subject. Is recommended using generic classes such as man, woman, or child (if the subject is a person) or cat or dog (if the subject is a pet).

Finetuning job

Finally, select number of saved checkpoints and click Start to start finetune Stable Diffusion on your dataset.

Finetuning job

The finetuning job usually takes about 2 hours for Stable Diffusion 2 and about 1.2 hours for Stable Diffusion 1.

3. Deploy finetuned model

After your finetuning job succeeded. Navigate through the Jobs side tab and expand to see your saved checkpoints. Click Save to choose the checkpoint which will be deployed.

Finetuning result

Navigate through the Deploy side tab and click Create deployments to deploy your saved models. Next, put your deployment name and select your saved model then click Deploy to deploy it.

Create deployment

Wait a bit until for deployment done then click to it to checkout how to use your deployment to generate images. We also provided example code to execute the API example for each programing language.

Deployment info

Note that your prompt should be start with sks prefix and include the class name you used in the finetuning phase. Below is the example with the prompt a photo of sks handsome football player raising the world cup championship trophy with img_height = 768 and img_width = 768.

Inference result