Tailored Generation

Description This route allows you to generate images using a Tailored Model. Tailored models are trained on a visual IP (illustrations, photos, vectors) to faithfully reproduce specific IP elements or guidelines. You can train an engine through our Console or implement training on your platform via API.

Overview of Tailored Generation Training API The Tailored Generation Training API provides a set of endpoints to manage the entire lifecycle of a tailored generation project:

1. Project Management: Create and manage projects which define the IP characteristics:

   • Create and Retrieve Projects: Use the /projects endpoints to create a new project or retrieve existing projects that belong to your organization.
   • Define IP Type: Specify the IP type (e.g., multi_object_set, defined_character, stylized_scene) and medium (currently illustration, with photography coming soon).
   • Manage Project Details: Use the /projects/{id} endpoints to update or delete specific projects.

2. Dataset Management: Create and manage datasets within your projects. You can:

   • Create and Retrieve Datasets: Use the /datasets endpoints to create new datasets or retrieve existing ones.
   • Upload and Manage Images: Use the /datasets/{dataset_id}/images endpoints to upload images and manage their captions.
   • Clone Datasets: Create variations of existing datasets using the clone functionality.

3. Model Management: Create and train models based on your datasets:

   • Create and Retrieve Models: Use the /models endpoints to create new models or list existing ones.
   • Choose Training Version: Select between "light" (for fast generation and structure reference compatibility) or "max" (for superior prompt alignment and enhanced learning capabilities).
   • Monitor and Control: Manage the model lifecycle including training start/stop and status monitoring.

Training Process Overview

To train a tailored model:

1. Create a Project: Use the /projects endpoint to create a project and define your IP type and medium.

2. Create a Dataset: Use the /datasets endpoint to create a dataset within your project.

3. Upload Images: Upload images to your dataset using the /datasets/{dataset_id}/images endpoint. Images must be 1024x1024px minimum.

4. Prepare Dataset: Review auto-generated captions and use the /datasets/{id} endpoint to set status to 'completed'.

5. Create Model: Use the /models endpoint to create a model, selecting light or max training version.

6. Start Training: Initiate training using the /models/{id}/start_training endpoint. Training takes 1-3 hours.

7. Monitor Progress: Check training status using the /models/{id} endpoint until status shows 'Completed'.

8. Generate Images: Once trained, your model can be used in multiple ways:

• Use the /text-to-image/tailored/{model_id} endpoint for text-to-image generation
• Use the /text-to-vector/tailored/{model_id} endpoint for generating illustrative vector graphics
• Use the /reimagine/tailored/{model_id} endpoint for generating using a structure reference images
• Access through the Bria platform interface

Alternatively, you can use the Bria platform to manage and train your tailored models through a user-friendly interface. Access the Bria platform here.

Guidance Methods

This API supports various guidance methods to provide greater control over text-to-image generation. These methods condition the model on additional inputs derived from user-provided images.

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.

Using ControlNets
You can specify up to four 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. The table below provides detailed information about each guidance method, with an example os use:

Guidance Method	Prompt	Scale	Input Image	Guidance Image	Output Image
ControlNet Canny	An exotic colorful shell on the beach	1.0
ControlNet Depth	A dog, exploring an alien planet	0.8
ControlNet Recoloring	A vibrant photo of a woman	1.00
ControlNet Color Grid	A dynamic fantasy illustration of an erupting volcano	0.7

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

• guidance_method_X: Specify the guidance method (where X is 1, 2). If the paramter guidance_method_2 is used, so does guidance_method_1 has to be used, and so on. If you would like to use only one method, use the paratmer 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

IP_adapter:

Guides the model based on the input image and its associated style. This method offers two modes:

• regular: Uses the full input image to condition the model, influencing both content and style.
• style_only: Focuses only on the style of the input image, allowing the model to generate images based on the provided aesthetic without altering the content.

Using IP-Adapter

Guidance Method	Prompt	Mode	Scale	Guidance Image	Output Image
IP-adapter	A drawing of a lion laid on a table.	regular	0.85
IP-adapter	A drawing of a bird.	style	1

Description

This route allows you to generate vector graphics using a Tailored Model. Tailored Models are trained on your visual IP (illustrations, photos, vectors) to preserve and faithfully reproduce specific IP elements or guidelines, ensuring consistency across all generated outputs. To see a detailed description of the tailored models' functionalities, please refer to the /text-to-image/tailored/{model_id} route documentation. *Text-to-vector is compatible with tailored models in the illustrative domain.

Guidance Methods

This API supports various guidance methods to provide greater control over text-to-image generation. These methods condition the model on additional inputs derived from user-provided images.

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.

Using ControlNets
You can specify up to four 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. The table below provides detailed information about each guidance method, with an example os use:

Guidance Method	Prompt	Scale	Input Image	Guidance Image	Output Image
ControlNet Canny	An exotic colorful shell on the beach	1.0
ControlNet Depth	A dog, exploring an alien planet	0.8
ControlNet Recoloring	A vibrant photo of a woman	1.00
ControlNet Color Grid	A dynamic fantasy illustration of an erupting volcano	0.7

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

• guidance_method_X: Specify the guidance method (where X is 1, 2). If the paramter guidance_method_2 is used, so does guidance_method_1 has to be used, and so on. If you would like to use only one method, use the paratmer 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

IP_adapter:

Guides the model based on the input image and its associated style. This method offers two modes:

• regular: Uses the full input image to condition the model, influencing both content and style.
• style_only: Focuses only on the style of the input image, allowing the model to generate images based on the provided aesthetic without altering the content.

Using IP-Adapter

Guidance Method	Prompt	Mode	Scale	Guidance Image	Output Image
IP-adapter	A drawing of a lion laid on a table.	regular	0.85
IP-adapter	A drawing of a bird.	style	1

Create a new project within the organization. A project encompasses all models trained and datasets created for the IP defined in the project.

The following IP types are supported:

Multi-Object Set A collection of different objects sharing a common style, design language, or color scheme. Objects are typically isolated on solid backgrounds.

Object Variants Multiple variations of the same object type, maintaining consistent style and structure while showing different interpretations. Objects are typically isolated on solid backgrounds.

Icons A collection of cohesive, small-scale illustrations or symbols designed to represent concepts, actions, or objects in interfaces and applications. Maintains consistent visual style across the set.

Defined Character A specific character that maintains consistent identity and unique traits while being reproduced in different poses, situations, and actions.

Character Variants Multiple characters sharing the same fundamental structure, style, and color palette, allowing creation of new characters that fit within the established design system.

Stylized Scene Complete environments or scenes created with a consistent visual style, look, and feel.

Retrieve all projects within the organization. If there are no projects, returns an empty array.

Retrieve full project information including project name and description, IP name and description, IP medium (photography/illustration), IP type, status, and timestamps.

Update a specific project

Permanently delete a project and all its associated resources, including all datasets, images, and models. This action cannot be undone. Training models must be stopped before deletion.

Create a new dataset.

Constraints:

• Dataset must have at least 1 image to be completed
• Maximum of 200 images per dataset

Retrieve a list of all datasets. If there are no datasets, returns an empty array.

Retrieve all datasets for a specific project

Retrieve a specific dataset

Update a dataset.

In order to use a dataset in a model training, its status must be set to completed.

Once a dataset status is changed to completed:

• Images cannot be added or removed
• Image captions cannot be edited
• Caption prefix cannot be modified

It is recommended to use the route Clone Dataset As Draft in order to create a new version of a dataset.

Constraints:

• Cannot update caption_prefix if dataset status is completed
• Dataset must have at least 1 image to be marked as completed

Delete a specific dataset. Deletes all associated images.

Create a new draft dataset based on existing one. This is useful when you would like to use the same dataset again for another training, but with some modification (create a variation).

Upload new image to a dataset.

Image Requirements:

• Minimum resolution: 1024x1024 pixels
• Supported formats: jpg, jpeg, png, webp
• Preferably use original high-quality assets
• For best results, use 1:1 aspect ratio or ensure main content is centered

Dataset Guidelines:

• Recommended: 5-50 images for optimal results when using Max training version, 15-100 for optimal results when using Light training version
• Maximum supported: 200 images
• Ensure consistency in style, structure, and visual elements
• Balance diversity in content (poses, scenes, objects) while maintaining consistency in key elements (style, colors, theme)
• Note: Larger datasets may introduce more variety, which can reduce overall consistency

For optimal training (especially for characters/objects):

• Subject should occupy most of the image area
• Minimize unnecessary margins around the subject
• Transparent backgrounds will be converted to black
• For character datasets: include diverse poses, environments, attires, and interactions

Captions and Generation:

• Each image receives an automatic caption that continues from the dataset's caption prefix
• Default caption prefix is recommended for initial training
• Captions can be modified to include domain-specific terms
• Both captions and prefix influence training and future generations
• Focus on essential elements rather than extensive details

Constraints:

• Dataset must have at least 1 image
• Dataset cannot exceed 200 images
• Cannot upload to a completed dataset

Retrieve all images in a specific dataset. If there are no images, returns an empty array.

Regenerate captions for all images in dataset. This action is crucial after the user updates the caption_prefix, and then it's recommended to regenerate all the captions of all images, to have full compatibility with the new caption_prefix.

This is an asynchronous operation. Once this endpoint is called, Get Dataset by ID should be sampled until the captions_update_status changes to 'completed'.

Retrieve full image information including caption (which naturally continues the dataset's caption_prefix), caption source (automatic/manual/unknown), image name, URL and thumbnail URL, dataset ID, and timestamps.

Update the caption of a specific image. There are two mutually exclusive ways to update a caption:

1. Provide a new caption text:

   • Use the caption parameter
   • This will set caption_source to "manual"
   • Reflects a human-written caption

2. Request automatic caption regeneration:

   • Set regenerate_caption to true
   • This will set caption_source to "automatic"
   • A new caption will be generated automatically based on the image and caption_prefix
   • For the same caption_prefix, regenerate_caption will always return the same caption
   • Useful for resetting captions or regenerating them after changing the caption_prefix

Note: You cannot provide both parameters simultaneously as they represent different update approaches.

Constraints:

• Cannot update captions in a completed dataset
• Cannot provide both caption and regenerate_caption in the same request

Permanently remove an image from a dataset. This will also delete the image files and associated thumbnails.

Constraints:

• Cannot delete images from completed datasets

Create new model. A dataset can be used to train multiple models with different training versions (e.g., one light and one max). The model will belong to the same project as its dataset.

Retrieve a list of models. If there are no models, an empty array is returned.

Retrieve all models for a project. If there are no models, an empty array is returned.

Retrieve full model information including name, description, status (Created/InProgress/Completed/Failed/Stopping/Stopped), training version (Light/Max), generation prefix, project ID, dataset ID, and timestamps.

Update a model's name and description. Other model attributes such as training version and dataset cannot be modified after creation.

Delete a specific model. Changes status to Deleted.