Best Practices for Tess Images

Tess can be an amazing tool for artists to brainstorm ideas, scale their freelance work, create seasonal and promotional materials, and make low-fi mockups for clients. You can also build a passive income stream if your Tess model is used by customers to generate editorial images. However, generating art in your style requires building an AI model that knows your visual style. 

This article will be a guide to selecting the best training set for Tess. If you’re choosing existing images to license, consider these guidelines.

The requirements for images on Tess are:

1. High resolution images: Images are at least 800 pixels in both dimensions. Any aspect ratio is fine. PGN or JPG file types.
2. Consistent visual style: Select images that have a consistent visual style. For example, line weight, color patterns, composition, level of detail, shapes, and medium all impact the visual style of a piece of art. When training Tess, upload images that match a coherent visual style for best results. 

3. Remove and avoid text. Avoid images with any typography, lettering, calligraphy, symbols, numbers, logos, or characters. Tess will attempt to redact or erase all lettering off of an image before training the model. Currency, our underlying AI model doesn’t ingest text well, so avoid using it in the training set.
4. Remove and avoid visual overlays like signatures, watermarks, or screenshot elements. If you have marks on the original images, the Tess model will learn these and impose them as blemishes in the outputs. Enable Tess to generate images in any aspect ratio by erasing or covering any visual marks in the original training set.

5. Full bleed: Use images flush to the outside edge. Tess performs best with scenes, still lifes, or illustrations that go all the way to the edge of the picture. Our AI model doesn’t perform as well with white or transparent backgrounds in the training set, and it doesn't perform as well for characters or icons.

6. Use single images rather than sets of illustrations: Sometimes, artists will combine images on a single canvas for their portfolio. These are common for logo artists, comic strips, and character sets. However, Tess performs best when each individual work of art is uploaded as its own submission in the AI training set.

7. Choose a diverse set of subjects. Tess works best with scenes, some of which include people and some of which do not, and learns the most with a diversity of different subjects. For example, a good training set might include a still life of a living room, two people shaking hands, a building on a city street, a bouquet of flowers, etc. Any subject matter works, with different perspectives, sentiments, and characters.

In addition to these requirements, you may also consider what visual style will help your Tess model stand out for our editorial customers. Share a model that expresses your creativity! The medium may help lifestyle bloggers, reporters on local news, radio station hosts, independent musicians, and content marketers generate visualizations for their writing, art, and work.

Tess represents a curated group of talented, experienced artists who have developed a style for editorial illustration. We have a high bar for quality, and our content moderators review every image that is submitted to Tess.