Freeform and Hi Res Freeform

What is it?

The Freeform process generates a sequence of images that are amalgamations of single collection. Through training, these resulting images resolve more clearly over time.

Using Alfred Sisley's landscapes as inspiration (left) to generate new landscapes (right).

What inputs are needed? What will I get from this process?

One collection of 30+ images is needed for the model, but a larger number, higher resolution, and more contextually similar images will create a better model.

Every snapshot will generate 50 new Freeform images. These images come at the 256 x 256px resolution in regular Freeform or at 1024 x 1024px resolution at Hi Res Freeform.

The 256px version of Freeform is good to prototype and get some results faster. The Hi Res Freeform takes more time and benefits from more input images, as it creates 1Mpixel images.

How do I set up the model?

From the Processes tab under Train a Model, choose Freeform, then click Create Project.

Freeform models require one input set of images: Inspiration.

You will need to Upload Images to make a new collection or choose a Recently Used collection, or add from Public Collections.

Only selected images will be used in the model.

You can Start Training your model by clicking the yellow button on the upper right corner.

  • The recommended training time for the 256px Freeform is 50 minutes, which roughly translates 50 snapshots with 50 images each, i.e. 2,500 resulting images.

  • The recommended time for Hi Res Freeform is 2.5 hours and you get 50 snapshots with 50 images each, i.e. 2,500 resulting images.

Once your model has begun training, you can access your Results within just a few minutes. The first images begin relatively noisy and resolve more clearly with more training.

Continue Training: Depending on your input images, some project might benefit from more iterations. You can click the orange Continue Training button in the top right to create more snapshots.

FAQs

How many images should I input?

The more the better! The minimum is 30, but you can upload as many as 5000. The images should be at least 256 x 256 px in resolution for the 256px Freeform and 1024px for the Hi Res Freeform. We accept PNGs and JPGs. Though uploading lower resolution images for your specific model is possible, results will likely exhibit a degraded quality.

What kinds of inputs gives the best results?

We find that a well-curated, large image set works really well. This image collection of Mark Zuckerberg photos is a good example. It has consistent cropping, and the eyes are "registered", meaning that in each image, the eyes are in the same place. However, the image collection would be better if it were higher resolution. The more consistent the images, the better the results.

How many results do you get?

You will get 50 resulting images per snapshot, regardless of the amount of images you inputted during training. For the Freeform model, you can also generate more resulting images using Mix, which uses the trained model from your most recent snapshot.

Many of my results look the same. Why?

This is called Mode Collapse, and it's a well-documented phenomenon that Generative Adversarial Networks (GANs) exhibit. Though it's a part of the mathematical properties of the AI model, we are looking into ways of decreasing its prevalence. This is more likely to happen if there are not enough variations in the inspiration images or if the model has been trained for very long time. Let us know if you find ways of decreasing the amount of mode collapse that you experience in your results. To learn more, check out this technical blog post: https://aiden.nibali.org/blog/2017-01-18-mode-collapse-gans/

What can I expect if I "continue training" on a Freeform project?

The results should look more and more similar to your input images as you train more snapshots/iterations. However, after a few hundred snapshots, any subsequent training will have very little effect on the results and can result in the images looking largely similar.

What about Wavy Freeform / Progressive?

You can read more about that here: