TikTok’s New AI Greenscreen Filter Turns Text into Images
TikTok has quietly debuted a new filter that creates background images from written suggestions using a text-to-image AI similar to OpenAI’s DALL-E model. The AI Greenscreen filter, while limited compared to what DALL-E can produce from the GPT-3 language model will likely introduce to a huge number of people the concept of text-to-image AI, thanks to TikTok’s global popularity. The AI Greenscreen may also still be under development as TikTok and its parent company ByteDance have yet to comment on it publicly. We’ve reached out to TikTok for comment.
TikTok AI Images
The AI Greenscreen filter can be found by searching for it like any other TikTok filter. The app offers a few suggestions for ideas to try but seems to limit them to a few adjectives and nouns. They all end up with a similar style of impressionist-style images that look much better with landscapes than with people or animals. In comparison, DALL-E and its rivals like Midjourney can create stunningly complex images in any specified style and with the realism depending mainly on the description. The simplified TikTok system is a good fit for the platform, however as the image is generated in just a few seconds and they are ideal as video backgrounds instead of having to stand on their own. It also helps skip over questions of what people would ask the AI to generate. If people are just swirls of color with a hint of a face there’s less for TikTok to worry about in terms of content moderation. Suggestions like ‘skydiving into rainbows’ and ‘stars above mountains’ make the style look like a deliberate choice.
The synthetic backgrounds may also pair well with TikTok’s new Avatars. Cartoonish custom digital representations of users built from template examples or from scratch. Users design skin color, hair texture, makeup, piercings, and a range of outfits, then pick avatar voice styles to bring the cartoon creation to life in future videos. The Miniature Avatar option will extend the virtual version of the user into the video as an independent creation that performs according to the content creator’s desires.