Stable Diffusion Gets an Upgrade—Here’s How Its AI Model Stacks Up Against MidJourney

AI startup Stability AI just debuted its latest version of Stable Diffusion—and the model does not disappoint.

Stable Diffusion XL (SDXL) v0.9 delivers ultra-photorealistic imagery, surpassing previous iterations in terms of sophistication and visual quality.

This means, among other things, that Stability AI’s new model will not generate those troublesome “spaghetti hands” so often. Also, you won’t have to introduce dozens of words to get an amazing image, because the model was trained to do most of the heavy lifting for you. Communicating with the model will be more natural. 

The company announced the release on Twitter yesterday, noting that the new version “provides a leap in use cases for generative AI imagery.”

Dubbed SDXL v0.9, the image generator excels in response to text-based prompts, demonstrating superior composition detail than its previous SDXL beta version, launched in April. A meticulous comparison of images generated by both versions highlights the distinctive edge of the latest model.

For instance, the prompt "A wolf in Yosemite National Park, chilly nature documentary film photography" rendered a more realistic image with the new AI model, outclassing the former version which fell short in its depiction of true-to-life details. Such significant enhancements are attributed to an increased count of parameters in SDXL v0.9, offering a greater depth of learning compared to its predecessor.

Stability AI, known for bringing the open-source image generator Stable Diffusion to the fore in August 2022, has further fueled its competition with OpenAI's Dall-E and MidJourney. Stable Diffusion is right now the world’s most popular open sourced AI image generator.

The company was recognized by TIME yesterday as one the most influential companies of 2023. Other AI companies that appear on the list are OpenAI (ChatGPT), Hugging Face (collaborative open source AI platform), Runway AI (generative video), Nvidia, and Google Deepmind. In the crypto space, Polygon and Chainalysis (blockchain forensics) also populate the list.

Beautiful Images With Less Work

In a notable shift, SDXL v0.9 ditches complicated prompts, generating better results from simpler, less structured inputs. This is vividly demonstrated when Decrypt submitted the short prompt "two hands pointing at each other bright art," yielding an impressively realistic result with SDXL v0.9, and less inspiring scribbles with the standard Stable Diffusion's versions 1.5 and 2.1.

This new ease of use could pose a serious threat to MidJourney, whose main appeal is its user-friendliness. Moreover, the cinematic aesthetics and precise object rendering by SDXL v0.9 could serve as a robust selling point for Stability AI, reminiscent of MidJourney's visual style.

Stability AI's latest gem will be accessible via Clipdrop, the AI image generating and editing tool developed by Init ML, a recent acquisition by Stability. The company's API customers are also due to gain access soon. However, the model is not yet ready for training or refining and doesn’t run locally. Once publicly released, it will require a system with at least 16GB of RAM and a GPU with 8GB of VRAM.

Meanwhile, Stability AI continues to develop the model alongside two other projects: a languishing large language model (LLM) named StableLM and the impressive DeepFloyd IF, an advanced text-to-image generator capable of embedding legible text into images—a feat not yet achieved by existing models.

According to Stability AI, mid-July is the anticipated date for the public release of this game-changing model as open-source software, marking another important milestone for the company.