How many images should my dataset contain?
RunwayML training experiments will accept individual datasets that total less than 5GB in combined file size and less than 25,000 images in total number. That's a lot, we know. Generally, the more data you collect, the better your model will be at generating synthetic images that look like they came from your original dataset, but a starting recommendation is to gather around 500-5,000 images. Some users, however, have been pleased with the results of a model trained on as few as 300 images.
What resolution is optimal for dataset images? What format?
The maximum resolution we currently support for dataset images is 178,956,970 total pixels (width x height). Images are resized to 320x320 as a preprocessing step during training, so having high resolution images should not have a significant effect on your model's performance. Runway accepts .jpg and .png files in dataset uploads.
Can I add more images to my dataset?
It is not possible to edit an uploaded dataset. If you’re interested in making changes to your dataset, they must be done locally and then re-uploaded to Runway.
Can I share my dataset with others?
No. Datasets are always private. They can only be viewed by the uploader and can never be used by anyone else.