ChatGPT’s 4o image generation model is the talk of the town right now, but it’s not the only AI software that can offer mind-blowing image generation. TikTok parent company ByteDance has a new AI model called InfiniteYou, whose sole purpose is to let users generate photos of themselves starting from a single uploaded photo.
It’s not that ChatGPT’s new image generation powers can’t edit photos you upload to the chatbot while preserving the identity of those characters. Other AI tools exist to let you edit your images in ways that fit your needs, even if that essentially means creating fakes; photos showing events that never happened and people who weren’t in that picture when it was taken.
However, the purpose of ByteDance’s new model is to generate fake pictures of a real subject while preserving their identity. That’s the whole point of InfiniteYou: To let you create any sort of image, starting from a simple photo upload that contains the main subject and a text prompt that describes what you want the AI to generate.
I’ll say from the get-go that the whole premise here is disturbing, not because I’m already worried about how incredibly easy it is to create lifelike fakes that can manipulate public opinion, but that the whole InfiniteYou research project comes from a company behind a product that’s often been accused of influencing public opinion via content algorithms. That’s social network TikTok, which still faces a major ban in the US.
The InfiniteYou service isn’t available as a standalone mobile app or web app, but you can test it at this link. Also, the AI research project is listed on HuggingFace, with the full study being available at this link.
As you’ll see in the following screenshot, you don’t even have to upload your own image to see what the AI can do. Just pick one of the available images, as I did, and then text a prompt. The photo I picked already came with the following prompt, so I didn’t even change it, as I was curious to see what the result would be:
A sophisticated gentleman exuding confidence. He is dressed in a 1990s brown plaid jacket with a high collar, paired with a dark grey turtleneck. His trousers are tailored and charcoal in color, complemented by a sleek leather belt. The background showcases an elegant library with bookshelves, a marble fireplace, and warm lighting, creating a refined and cozy atmosphere. His relaxed posture and casual hand-in-pocket stance add to his composed and stylish demeanor
The AI took a while to process the uploaded photo and the requirements in the text, and then it generated the image on the right side here:
As you can see, the AI image preserved the subject’s likeness and recreated the entire background and the subject’s body to adhere to the prompt’s requirements.
The resemblance between the subject in the photo and the AI version is clear, though you can tell the image on the right is AI-generated. There’s no watermark to indicate it’s an AI photo (which itself is a red flag), but you can tell this isn’t a real photo.
Perhaps that’s a good thing. Otherwise, InfiniteYou could be easily used to create deepfakes of celebrities in lifelike photos, a problem the new ChatGPT image generation model already has.
Then again, I only briefly tested this new AI on HuggingFace. A commercial product will likely offer even higher-quality images that are harder to identify as AI-generated images.
After all, the images the researchers offered in the study suggest that the AI model can create high-quality, albeit fake, images of a subject with the help of a real photo and a text prompt.
Take the examples above, each containing the original photo, the text prompt InfiniteYou was given, and the result. We are looking at high-end, frontier AI tech here.
The ByteDance engineers also provided the following comparison between InfiniteYou and other AI models that can generate images.
It’s unclear where ByteDance might use this AI tech next, but it’s clear where it might want to deploy it. TikTok comes to mind again, as AI tech like InfiniteYou would certainly come in handy to creators.
That would not be a problem as long as AI content is clearly labeled as such and is not used for malicious purposes.
The AI researchers addressed safety concerns in the study, but only briefly. Rather than offering solutions to prevent fakes, they suggest InfiniteYou can be further improved. As for creating fake images, the researchers say they “developing robust media forensics approaches can serve as effective safeguards:”
Limitations and societal impact. Despite promising results, the identity similarity and overall quality of InfU could be further improved. Potential solutions include additional model scaling and an enhanced InfuseNet design. On another note, InfU may raise concerns about its potential to facilitate high-quality fake media synthesis. However, we believe that developing robust media forensics approaches can serve as effective safeguards.
Who will develop those safeguards? Who knows?
Meanwhile, you can explore ByteDance’s sophisticated InfiniteYou AI model at this link.
