Articles on: Model customization
This article is also available in:

Model training overview

Seelab.ai allows you to train your own models, based on your style, a character, a photography style, or even a product. This will enable you to generate images with very high consistency as you continue generating.

Video Tutorial





Accessing Model Training



To access the training, click on the "custom models" tab in the menu.



You will land on a page where you can find all the models you have previously created.

If this is the first model you're creating, you won't see this page, you'll directly access the "training type selection" section.

Hovering over the model card, you will be able to:

View the images created with this model
Select it to create images
Delete it



Warning, deleting a model is permanent. You won't be able to recover it later.

To create a new model, click on the button ... "create a new model"

Choosing the Training Type



Seelab currently offers 4 different types of training



Avatars and Characters: Transform yourself into anything you want and create stunning avatars or train a character representing your brand, or even a pet :)
Style and Illustrations: Train a model in the artistic style of your brand guidelines and create visuals that match your brand in just seconds
Photography: Generate a model that will produce realistic and coherent images while maintaining lighting, staging, style, and more!
Produit: Generate a model that will produce your product in any situation


More training types are coming soon, including training models for logos/icons.

Choose a name for your model: try to pick a short and understandable name, it will appear in the "creation" section to allow you to select it.

Once you’ve chosen the training template and edited the name, click on the "next step" button at the top right of the interface.

Choosing the Dataset



Uploading the Dataset and Briefing the AI





A successful model starts with a good dataset! Make sure you've previously selected between 10 and 25 images that best represent your style or the character you wish to replicate.

For more information on how to build a good dataset, check out the article "The secrets to a successful model."

Click on the "upload images" box and select visuals from your computer.

Some rules to follow:

We require a minimum of 10 images to train a model
Uploaded images must have a resolution of at least 1024 pixels
They must not be similar (no duplicate images)

Selecting the Most Representative Images (optional)



You can select up to 3 images that best represent the style or character you want to replicate.
This will help Seelab's AI refine the model. To do this, hover over an image and click the "thumbs up" icon.



Writing Your Brief (optional)



You can also brief the AI by explaining what you want to achieve with this model:

Provide details on elements that should be retained and preserved
Conversely, specify if certain points should be disregarded or less emphasized
Feel free to describe in your own words how you define this style
Add the context in which you'll use the images

This information will be analyzed by a text model and will influence the model and its quality.



Training the Model



When you’re ready, click the "start training" button.

The training process is complex and requires powerful resources. That’s why you won’t get the result right away.

You will be notified by email (the one used during registration) when the training is complete. While the process usually takes less than an hour, it may take several hours depending on activity on the Seelab platform. We guarantee a result within 24 hours.

Accessing Your Trained Model and Generating Images



Once you receive a notification by email, you will be able to find it in the "creation" section, under the "project models" tab.



Click on it to start generating stunning images! If you've invited collaborators to a project, they will also be able to access your model.

It's possible that our system may refuse to train a model if it doesn't comply with our terms of use or if the style coherence or image resolution criteria are not met. In that case, you'll need to revise your dataset. If so, don't hesitate to contact us!

Updated on: 05/11/2024

Was this article helpful?

Share your feedback

Cancel

Thank you!