| Prisma Labs | |
| Industry: | Artificial intelligence |
| Hq Location: | Sunnyvale, California, US |
| Products: | Prisma, Lensa |
Prisma Labs is a software company based in Sunnyvale, California that is known for developing Prisma and Lensa.[1]
Prisma Labs was founded in 2016 by Andrey Usoltsev, Alexey Moiseenkov, and a team of Russian developers.[2] [3] Usoltsev is also the CEO. In 2016, the company launched the Prisma app, which uses artificial intelligence to duplicate photos in various artistic styles. In 2018, the company launched the Lensa AI app, which is a photo and video editing app. In late November 2022, Lensa's "magic avatars" feature was launched, which, for a fee, uses artificial intelligence and users' uploaded selfies to create portraits of the users in various styles and settings within minutes.[4] Lensa uses Stable Diffusion, an open source text-to-image model launched by Stability AI in August 2022. The company says it uses user photos to train its AI, and its user agreement states that Lensa can use the photos, videos, and other user content for "operating or improving Lensa" without compensation.[5] The Lensa app has been criticized for producing hypersexualized images of women and girls, including non-consensual pornographic content, a bias not present when processing images of men.[6]
Prisma Labs uses the Stable Diffusion generative engine to power the Lensa apps’s Magic Avatar feature. [7] Stable Diffusion is open source and was trained on the LAION 5B dataset,[8] which utilized 5.85 billion CLIP-filtered image-text pairs from Common Crawl to create the dataset. [9]
There are two main areas of contention with regards to the LAION 5B dataset. The first is the ownership of the images it contains, which were largely obtained from public sites like Pinterest. [10] The copyright of those images is generally held by people and companies that have nothing to do with any kind of dataset or AI.[11] There is currently a class-action lawsuit happening against companies that use Stable Diffusion on the grounds of copyright infringement.[12]
The second area of contention is the biases inherent in the images contained in the dataset. There are allegations of Lensa/Stable Diffusion demonstrating racist stereotypes,[13] fatphobia,[14] and white-centredness.[15] There are also reports of the engine creating nude or suggestive images[16] without being prompted to. Since the release of 5B, LAION has implemented a NSFW filter, which Stable Diffusion has incorporated, but LAION states that it still cannot guarantee the results. Lensa has issued the following statement about the issue: “The Stable Diffusion model was trained on unfiltered Internet content. So it reflects the biases humans incorporate into the images they produce.”[17]
Prisma Labs acknowledges the biases in the datasets, claiming that the feature “reflects the biases humans incorporate into the images they produce.”[18]