Train your own model
Train a custom find-tuned model for your unique needs. Everything you train remains strictly private unless you publish and share with the community.
Tutorial Video
Style model
- Quantity matters a lot. The more images you curate, the better AI can capture the style.
- Consistency is key. There should be a clear common theme within the uploaded images.
- Variety wins. If the training images only contained cars, the model will have a hard time generating humans - unless you only want car images back.
Try the “Complex Character Designs” model →
Character model
- Have a variety of poses, angles and background.
- Try to limit the appearance of characters other than the one you’re training for.
- Avoid images that lack stand-out features that are vital to your character. Differences between images make the model think that those are changeable features.
- Mix as many media formats as possible: screenshots of 3D renders, pictures from 3D prints, “tradigital” painting, etc.
Face model
- Similar to training a character model, have a good variety of outfits, hairstyles, backgrounds, facial expressions, angles and poses.
- Use photos with a single well-lit human subject without obstructions like glasses or masks.
- Close-up or half-body images are preferred over full-body images.
Object model
- Have photos of the object from different angles and different distances.
- Use a single clear English word for the object prompt box, like “car” “mug” “cat” - don’t use your cat’s name “Lucy”. We’re trying to tell AI what this object is by nature.
- Objects should preferably stand out against the background.
Quick tips
- The most essential thing about training a model is the images you upload.
- Square images are preferred. Images are cropped to square during training.
- The higher the training image quality, the higher the output image quality.
- It may take more than one iteration to obtain a great model. You will observe the results, re-curate the dataset accordingly, and upload the dataset train a new model.
- Note that you cannot directly edit the dataset of a trained model to re-train it. This helps keep your data private and secure.
The Model Training Book
By model training, we mean “fine-tuning” here. This allows AI to learn a concept from the images you provide and generate images from that concept. The Model Training Book is brought to you by OpenArt, with concrete examples covering use cases from character consistency, to your unique style as an artist, to product/game/architecture/interior/jewelry designs, to pet lovers.
Sharing your trained model
Publish your model so others can create with it from Community Models.