Overview

Training Process

To train a tailored model:

1. Create a Project: Use the /projects endpoint to define your IP type and medium.
2. Create a Dataset: Use the /datasets endpoint to create a dataset within your project.
3. Generate an Advanced Caption Prefix (For stylized_scene, 'defined_character' and 'object_variants' IP types):

• Before uploading images, call /tailored-gen/generate_prefix, sampling 1-6 images from the input images provided for training (preferably 6 if available).
• Update the dataset with the generated prefix using /datasets/{dataset_id}.

3. Upload Images: Upload images using the /datasets/{dataset_id}/images endpoint (minimum resolution: 1024x1024px).
4. Prepare Dataset: Review auto-generated captions and update the dataset status to 'completed'.
5. Create Model: Use the /models endpoint to create a model, selecting either the "light" or "max" training version.
6. Start Training: Initiate training via the /models/{id}/start_training endpoint. Training typically takes 1-3 hours.
7. Monitor Progress: Check the training status using the /models/{id} endpoint until training is 'Completed'.
8. Generate Images: Once trained, your model can be used in multiple ways:

• Use /text-to-image/tailored/{model_id} for text-to-image generation.
• Use /text-to-vector/tailored/{model_id} for generating illustrative vector graphics.
• Use /reimagine/tailored/{model_id} for structure-based generation.
• Access through the Bria platform interface.

Alternatively, manage and train tailored models through Bria's user-friendly Console.
Get started here.

Reimagine - Structure Reference

The Reimagine - Structure Reference feature lets you guide outputs using both a tailored model and a structure reference image.

It produces visuals that preserve the structure of the reference image while applying specific characteristics defined by your tailored model. Access this capability via the /reimagine endpoint.

Restyle Portraits

The Restyle Portrait feature enables you to change the style of portrait images while preserving the identity of the subject. It utilizes a reference image of the person alongside a trained tailored model, ensuring consistent representation of facial features and identity across different visual styles.

Restyle Portrait is specifically optimized for portraits captured from the torso upward, with an ideal face resolution of at least 500×500 pixels. Using images below these recommended dimensions may lead to inconsistent or unsatisfactory results.

To use Restyle Portrait:

• Reference Image: Provide a clear portrait image that meets the recommended guidelines.

• Tailored Model: Select a tailored model trained specifically to reflect your desired style or visual IP.

Use the /tailored-gen/restyle_portrait endpoint to access this capability directly, allowing seamless integration of personalized style transformations into your workflow.

Guidance Methods

Some of the APIs below support various guidance methods to provide greater control over generation. These methods enable to guide the generation using not only a textual prompt, but also visuals.

The following APIs support guidance methods:

• /text-to-image/tailored
• /text-to-vector/tailored

ControlNets:
A set of methods that allow conditioning the model on additional inputs, providing detailed control over image generation.

• controlnet_canny: Uses edge information from the input image to guide generation based on structural outlines.
• controlnet_depth: Derives depth information to influence spatial arrangement in the generated image.
• controlnet_recoloring: Uses a grayscale version of the input image to guide recoloring while preserving geometry.
• controlnet_color_grid: Extracts a 16x16 color grid from the input image to guide the color scheme of the generated image.

You can specify up to two ControlNet guidance methods in a single request. Each method requires an accompanying image and a scale parameter to determine its impact on the generation inference.

When using multiple ControlNets, all input images must have the same aspect ratio, which will determine the aspect ratio of the generated results.

To use ControlNets, include the following parameters in your request:

• guidance_method_X: Specify the guidance method (where X is 1, 2). If the parameter guidance_method_2 is used, guidance_method_1 must also be used. If you want to use only one method, use guidance_method_1.
• guidance_method_X_scale: Set the impact of the guidance (0.0 to 1.0).
• guidance_method_X_image_file: Provide the base64-encoded input image.

Image Prompt Adapter:

This method offers two modes:

• regular: Uses the image’s content, style elements, and color palette to guide generation.
• style_only: Uses the image’s high-level style elements and color palette to influence the generated output.

To use Image Prompt Adapter as guidance, include the following parameters in your request:

• image_prompt_mode: Specify how the input image influences the generation.
• image_prompt_scale: Set the impact of the provided image on the generated result (0.0 to 1.0).
• image_prompt_file: Provide the base64-encoded image file to be used as guidance.

or

• image_prompt_urls: Provide a list of URLs pointing to publicly accessible images to be used as guidance.

IP-related prompts

Our models are trained exclusively on fully licensed, safe-for-commercial-use data.
Prompts that reference public figures, brands, or other protected content may result in generic or altered outputs.
These prompts are not blocked, but results may differ from what you expect.
If an IP-related signal is detected in the prompt, the following warning will appear in the API response: