AI Photoshoot (DreamBooth)
Train SD to make images and artworks of any person or object
Works for people, pets and objects
Realistic and artistic images
Minimum 3 photos needed
How to work with it
1. Upload images
If you train on a person, we recommend to use portrait images. Cut your images to 512x512 for better results.
2. Connect images
3. Describe your images
The most important part is to name right two fields: 1) Class prompt. Use 'a photo of a man, ultra detailed' or 'a photo of a woman, ultra detailed' for best results 2) Subject. Use a unique name of a person (avoid simple names and blank spaces) and add 'a man' or 'a woman' after it
4. Launch and wait
Don't get scared, the images will be better in Stable Diffusion!
5. Use in Stable Diffusion
Parameters for training
Subject
It should describe what is shown on your uploaded images. Later you will use what you write in this text field as a prompt in Stable Diffusion. It has to be unique, something that Stable Diffusion doesn't know yet. It must contain two things: 1) unique name of person. Avoid simple names (use DariaWind, not just Daria) and blank spaces (DariaWind, not Daria Wind) for better results 2) subject type (man, woman, shoe, etc.) For example, 'DariaWind, a woman', 'EkonikaBoot, a shoe'
Steps
We recommend to use around 1200-1500 steps. It will take longer (around 30 min), but you will get more similarities to original person.
Class prompt
This prompt is used by AI to generate additional images. These additional images will make the training results better. These prompts work great for training on people: "a photo of a woman, ultra detailed" "a photo of a man, ultra detailed"
Training on objects
You can train model on any object in Phygital+. Follow these recommendations for the best results
How many photos?
We recommend using at least 5 shots from different angles
Steps
The recommended number: 1500
Do I need Class images and prompt?
Yes, we adivse to set these parameters:
Class images: 10
Class prompt:
' a photo of a
What do I put into Subject?
The naming follows the same rule as with people and contains two parts: 1) unique name (for example, GucciDianaTote) 2) subject type (for example, a bag) So, the full Subject type will be 'GucciDianaTote, a bag' for the pictures on the left
Training on style
You can a train model on any object in Phygital+. Follow these recommendations for the best results
How many photos?
We recommend using at least 20 refrences, but the more, the better
Steps
The minimum number: 1500
Do I need Class images and prompt?
Generally no, but in our interface we adivse to set these parameters:
Class images: 10
Class prompt:
' style of
What do I put into Subject?
The naming follows the same general rule and should have: 1) a unique name (something that Stable Diffusion doesn't know yet) 2) phrase 'in the style' So, the full Subject will be 'in the style of ProjectWinter game' for the pictures on the right
