Model Training (ModelGenerator)

Fine-tune a base model on your custom dataset. ModelGenerator is the backbone for personalized avatars, advanced product photography, domain-specific image generation, and anything else where you need a model that “knows” your subject.

This is usually the step where you’ll spend the most time tinkering — small hyperparameter changes can make a big difference in output quality. You can also train models directly through the dashboard and generate images from them there.

Preparing a dataset

High-quality data is the single most important factor in training. For example, if you’re training a model to learn a person’s appearance, all images should have good lighting and resolution, include different angles and poses, and contain only one person.

Since you don’t control uploaded datasets for public Flows, instruct your users on how to best select their images. There are also hyperparameters you can configure to minimize the impact of a suboptimal dataset — but you’ll need to experiment.

Upload via Dashboard:

• Go to the Subjects section.
• Create or select a subject and upload your images (.jpg or .png).

Upload via Flow Input:

• Declare an input property with "dataset": true (e.g., "userDatasetId").
• When a user runs the Flow, they’ll be prompted to upload files.
• These are automatically assembled into a dataset under your organization.

Each dataset has a unique id you can reference in the training parameters.

Step definition

In your Flow’s steps array, add a ModelGenerator step:

{
  id: 'trainer',
  type: 'ModelGenerator',
  parameters: {
    datasetId: '$.input.userDatasetId',
    settings: {
      baseModel: 'stable-diffusion-v1-5/stable-diffusion-v1-5',
      epochs: 50,
      resolution: 1024,
      normalizeDataset: true,
      maskedLoss: true,
      // ...more hyperparameters below
    },
  },
}

Parameters

Hyperparameters

Core

Dataset processing

Optimizer

Network

Advanced

Example

Here’s a real-world training configuration for portrait training:

{
  id: 'portraitTrainer',
  type: 'ModelGenerator',
  parameters: {
    datasetId: '$.input.userDatasetId',
    settings: {
      baseModel: 'Realistic_Vision_V6.0_B1',
      epochs: 80,
      repeats: 1,
      alphaDim: 64,
      decouple: true,
      batchSize: 1,
      optimizer: 'Prodigy',
      precision: 'fp16',
      scheduler: 'Cosine',
      maskedLoss: true,
      networkDim: 64,
      resolution: 1024,
      learningRate: 1,
      bodyMaskShade: 200,
      headMaskShade: 255,
      biasCorrection: true,
      normalizeDataset: true,
      boundingMaskShade: 40,
      backgroundMaskShade: 40,
    },
  },
}

Training templates

Repeatedly specifying the same training parameters can be tedious. Training Templates let you define a configuration once and reuse it.

1. Go to Trainings → New training template in the dashboard.
2. Fill out fields like base model, epochs, etc.
3. Save it as a template.
4. Reference it in a Flow:

{
  id: 'trainer',
  type: 'ModelGenerator',
  parameters: {
    trainingSettingsId: 'your-template-uuid',
    datasetId: '$.input.datasetId',
  },
}

You can still override or supplement the template’s settings in your Flow’s settings block.

Best practices

• Dataset quality matters most — ensure images are clear, well-lit, diverse in angles/poses, and properly represent your subject.
• Start with defaults — define only a small subset of hyperparameters at first. As results improve, experiment with more.
• Use a fixed seed when testing — this keeps other variables constant so you can isolate the effect of each change.
• Some hyperparameters work in combination — for example, maskedLoss works best with normalizeDataset: true, and Prodigy benefits from biasCorrection and decouple.
• Verify your base model — it must be a valid Hugging Face repository name. If you see errors, double-check the spelling.