AI imaging service Midjourney has begun alpha testing of the fourth version (v4) of its text translation model to an image that is available to subscribers on the Discord server, reports ArsTechnica. The new model provides greater detail than previously available version, which inspired some AI artists to note that v4 almost makes “too easy” to get high-quality results with simple tips.
Midjourney opened to the public in March as part of an early wave of AI-powered imaging models. It quickly gained a large following thanks to its expressive style and the fact that it became publicly available before DALL-E and Stable Diffusion.
Comparison of Midjourney v3 (left) and v4 (right) with the hint “muscular barbarian with weapon near CRT TV, cinematic, 8K, studio lighting”.
Over time, Midjourney refined its model with more training, new features, and more detail. The current default model, known as “v3”, was introduced in August. Midjourney v4 is currently being tested by thousands of members on the service’s Discord server, who create images using the Midjourney bot. Users can currently try v4 by adding “–v 4” to their prompts.
“V4 is an all-new code base and all-new AI architecture,” Midjourney founder David Holtz wrote in a Discord announcement. “This is our first model trained on the new Midjourney AI supercluster and has been running for over 9 months.”
When testing the Midjourney v4 model, it was found to provide much more detail than v3, better understanding of hints, better scene composition, and sometimes better object proportions. If you’re aiming for photo-realistic images, some results may be hard to tell from real, lower-resolution photos.
According to Holtz, v4’s other features include:
- Much more knowledge (about creatures, places, etc.);
- Recognizes small details much better (in all situations);
- Handles more complex prompts (with multiple levels of detail);
- Works better with multi-object / multi-character scenes;
- Supports advanced functionality such as image hints and multitips;
- Supports –chaos arg (set it between 0 and 100) to control the variety of image grids.
Midjourney states that v4 is still in alpha, so developers will continue to work on the new model. The company plans to increase the resolution and quality of v4 zoomed images, add custom aspect ratios, improve image clarity, and reduce text artifacts. Midjourney is available for a monthly subscription fee that ranges from $10 to $50 per month.