Content
summary Summary

Playground, a team of developers specializing in graphical AI, has released Playground v2, a new diffusion-based generative text-image model.

Ad

According to a user study conducted by Playground, images generated with Playground v2 are preferred 2.5 times more than images generated with Stable Diffusion XL. The team ran a comparative study in which thousands of users were asked to choose their favorite image between the two models.

Image: Playground

Playground offers its new model as a basic version with resolutions of 512px and 256px at Huggingface. An "aesthetic" version of the model with a resolution of 1024px is also available for download, as is the v1 version.

Although the resolution of the open models is relatively low, the AI-generated images can be easily scaled up using AI upscalers without significant loss of quality.

Ad
Ad

The AI image model is also available on the Playground website. Anyone who creates an account here can generate 500 images per day for free. The images can be used commercially.

For twelve US dollars per month, users have access to numerous image editing and prompting functions via a graphical user interface. Starting at one million images per month, Playground also offers API access.

Image: Playground AI

Playground AI was founded in San Francisco in 2022 and has raised about $40 million from investors to date. Its stated goal is to create visual AI systems that can make subtle changes to images based on prompts, including the creation of entire 3D environments or complex analysis of video scenes. The image generator is just the first step.

New benchmark and license conditions

In addition to the new model, Playground has also introduced a benchmark, MJHQ-30K, to automatically evaluate the aesthetic quality of a model. The benchmark calculates the Frechet Inception Distance (FID) based on a high-quality dataset. The dataset, curated with Midjourney, contains ten general categories with 3,000 examples each. It also shows how well the image AI can follow prompts.

Image: Playground AI

Playground v2 is available for download from Huggingface under a "community license". You are granted a non-exclusive, worldwide, non-transferable, royalty-free, limited license under Playground's intellectual property rights to use, reproduce, distribute, copy, create derivative works from, and modify the Playground material. This license is free for research and commercial use.

Recommendation

However, the license is subject to certain limitations. For example, if the creation or manipulation of images is a primary business or product of Licensee and the total number of monthly unique users ("MUUs") of the products or services provided by Licensee or its affiliates for or on behalf of such products or services exceeds one million MUUs in the preceding calendar month, Licensee must apply for a license from Playground.

Ad
Ad
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.
Support our independent, free-access reporting. Any contribution helps and secures our future. Support now:
Bank transfer
Summary
  • Playground, a team specializing in graphical AI, has released Playground v2, a diffusion-based text-image model that was 2.5 times more preferred than Stable Diffusion XL in user studies.
  • The base version of the model is available for download at Huggingface in resolutions of 512px and 256px, in addition to an "aesthetic" version of the model with a resolution of 1024px. Users can create 500 commercially usable images per day on the Playground site.
  • Playground has also introduced a new benchmark, MJHQ-30K, which allows the automatic evaluation of the aesthetic quality of a model by calculating the Frechet Inception Distance (FID) on a high-quality dataset.
Online journalist Matthias is the co-founder and publisher of THE DECODER. He believes that artificial intelligence will fundamentally change the relationship between humans and computers.
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.