Model Training

Reference Images are uploaded then deleted off our system. The result - Arcana can now learn any style, concept or subject you wish to create. The sky is the limit.

woman with curly hair and golden dress among yogis

HOW TO TRAIN

1. Inside the Model Training Wizard make sure to have high quality images ready.

2. For best results, focus on a SINGLE CONCEPT such as a CHARACTER or GENERAL STYLE. PORTRAITS: Neutral backgrounds. Females, please use limited makeup. Focus on skin texture to improve results. (7-15 images)

PRODUCTS: Less is more. (5-10 images)

CINEMATIC: Use high resolution reference images from movies with a style you want to emulate. (20-100 images)

CREATURE DESIGN: Blend different creature images together for cool experimental designs. (3-20 images)

MISC: You can teach Arcana ANYTHING! Upload pictures of interior designs you like, ceramic art you want to recreate, famous paintings you want to style emulate. THE SKY IS THE LIMIT!

1. After submission and a brief training period, it's time to test your new model! Adjust your new model's guidance levels in the advanced tab between 2.5-3.5 to find the golden range. Default is 3.5.

BOTTOM LEFT: Adjust your guidance levels.

1. If you provided a good dataset, your prompts should look excellent.

TRAINING FOR CREATURE DESIGN

Look at the images in your dataset that you use to train as ingredients in a milkshake. The key is BLENDING ingredients. The more of the same images you use, the more of a bias, or pronounced flavor that will emerge in your dataset. If you only train on 3 images a blend can occur where you have ⅓ of the model focused on each of the images in your dataset.

This is awesome for Creature design, just look here for an example:

Right: Dataset images used to train. Only 3 images! The Left: AI image, fully original creature design that blended the 3 ingredeients of our dataset

AI image made off dataset with gross creatures. This a wholly original creation made through prompting.

AI image made off dataset with gross creatures. This a wholly original creation made through prompting.

TRAINING FOR WORLD BUILDING or GENERALIZED AESTHETIC

Continuing with our milkshake metaphor, the better the ingredients, the better the product. Thus, the higher the quality images in your dataset, the higher the quality model and resulting images.

When creating a style-based, or world-based model more volume is preferable - around 20-100 images.

For beautiful cinematic worlds, upload stunning HD images of cinematic shots you want to homage. Look at it as ARCANA AI learning your mood board.

REMEMBER: ONLY UPLOAD images you have the rights to!

Ai image generated from custom model trained on beautiful photographs

AI image made after trained on beautiful dataset of exotic properties.

TRAINING FOR PORTRAITS AND CHARACTER CONSISTENCY

With as little as 5 images, you never need a camera again!

When training self-portraits, make sure that you use neutral backgrounds so there isn't a bias on the background built into the training model.

For women, be consistent with the makeup you're wearing and try and upload images with as much capturable skin texture as possible.

Once trained, you should be able to take endless photoshoots of yourself and NEVER USE A CAMERA AGAIN!

Left: REAL SELFIES uploaded. Right: Ai generated images in a variety of settings

AI image made after training on a man's selfies. No images involved war or war gear in the dataset. It was generated solely through prompt with brilliant character consistency.

FURTHER WORKFLOWS

Using a Character Trained Model for cast consistency with inPaint:

Click here for step-step instruction