Unlocking New Frontiers in AI Image Creation
Written on
Midjourney 5.2: A Transformative Update with New Features
Midjourney has just rolled out its latest update, and it's a significant leap forward in AI image creation. The introduction of outpainting is a highlight of this update, marking a substantial enhancement!
Beyond this, the platform is making strides to keep pace with the latest advancements in AI image generation, presenting a host of exciting capabilities. The new features include:
- Outpainting (finally!)
- Enhanced variation modes
- Automatic prompt optimization
- Notable improvements in quality and coherence
- And more!
New Aesthetic Framework
Quality & Coherence
Let’s examine the distinctions between the new V5.2 model and its predecessor, V5.1. For this comparison, we utilized the seed parameter (--seed 1000). Here’s the prompt we employed:
cinematic shot, astronaut in the jungle --seed 1000
We will incorporate both --v 5.1 and --v 5.2 to call the respective Midjourney model versions.
Here are the outcomes:
The results already reveal an improved coherence and enhanced visual quality.
These improvements will be even more pronounced when used alongside other updates like the --stylize command, outpainting, and variation ranges.
Stylize Command Exploration
To investigate the updates within the new --stylize command, we applied this prompt:
cinematic shot, astronaut in the jungle --seed 1000 --v {5.1, 5.2} --stylize {0, 50, 200, 600, 1000}
For those unfamiliar, the bracket notation serves as Midjourney's permutation feature, allowing multiple prompts to be processed simultaneously.
Results for V5.1:
As anticipated, the --stylize command has minimal impact at values exceeding 200, as the effect range has been intentionally limited since version 4.
Results for V5.2:
The --stylize command is now fully functional in version 5.2, capable of altering the image up to the maximum value of 1000.
New Outpainting Feature
After “upscaling” an image via the “U” button, users can now select from four outpainting options:
- 2x zoom
- 1.5x zoom
- “Make Square” option
- “Custom Zoom” option
Here are examples of the 1.5x and 2x zooms:
The custom zoom feature allows users to define a zoom level between 1.0 and 2.0, along with a tailored aspect ratio. For instance, using a 1.8x zoom and a cinematic widescreen aspect ratio of 21:9 yields:
The “Make Square” option converts any image into a 1:1 aspect ratio using outpainting.
These features provide incredible new reframing options, which I plan to explore in greater detail in my upcoming series on cinematic prompts.
Additionally, you can manipulate the original prompt while executing custom zooms, leading to fascinating alterations in the image during outpainting. Here's an example after several zooms:
New Variation Modes
Exclusively for V5.2, you can switch the variation mode by typing /settings and selecting either “High Variation Mode” or “Low Variation Mode”:
Let’s investigate how this operates.
From the initial grid (left image), I upscaled image #3 (right image):
After upscaling, two new options appear in V5.2:
Here’s the comparison of “Vary (Strong)” and “Vary (Subtle)” with “High Variation Mode” enabled:
Now, observe “Vary (Strong)” and “Vary (Subtle)” in “Low Variation Mode”:
The differences are subtle across all versions due to the narrow nature of the prompt, limiting variation. A broader prompt yields more noticeable effects:
Now, with “High Variation Mode” enabled, let’s rerun “Vary (Strong)” and “Vary (Subtle)” to observe increased variation in the outputs:
New /shorten Command
The new /shorten command offers a valuable tool for analyzing prompts, identifying key and less significant words.
The syntax is: /shorten + PROMPT
Upon submitting your prompt, you receive a standard response along with an optional detailed view showcasing the bot’s suggestions (note: /shorten does not support multi-prompts).
Note that these are merely suggestions.
The bot may not always accurately assess the significance of words, as illustrated in the above examples regarding recommended shortened prompts.
Moreover, the evaluation of crucial tokens (elements of a prompt understood by the model) isn’t flawless. Minor modifications to words deemed unimportant by the bot can significantly steer the model's direction.
Here’s an example with V5.2:
- left prompt: cinematic shot, astronaut in the jungle — seed 1000
- middle prompt: cinematic, astronaut in the jungle — seed 1000 — v 5.2
- right prompt: astronaut in the jungle — seed 1000 — v 5.2
Despite its imperfections, this feature clearly illustrates the evolving nature of “prompt engineering”:
AI algorithms are becoming more adept at analyzing our instructions and suggesting improvements, steering us closer to automatic prompt optimization.
Looking Ahead
- As per their Discord announcement, changes may occur in the upcoming weeks without prior notice since V5.2 is still in testing.
- To gather maximum feedback, V5.2 is the default setting, but users can revert to V5.1 at any time by typing /settings and selecting V5.1.
- On Discord, check out the community showcase and discuss new results in the ideas-and-features and show-and-tell channels.
- High Variation Mode is exclusive to V5.2.
Stay connected with me for updates on "AI & Creativity." If you wish to support my work, consider becoming a Medium member through my referral link, granting you full access to all my articles (140+ and growing) and those of countless other writers.
If you enjoy my content, please leave a “clap” at the end of this article, so more people can discover it!