Quick Midjourney Update: 4 key highlights to know (21st December 2023)
Keeping up with the rapid changes and developments
Merry Christmas and Happy Holidays!
Here’s a quick rundown of the most recent Midjourney Weekly Office Hours:
(1) Upcoming releases
- Midjourney V6 Alpha may be released by today or tomorrow. However, if anything unexpected occurs, the date may be pushed back to January.
- Website features — see below.
(2) Midjourney Website
- The new folder or collection feature will be available next week.
- The group area feature may be released in January 2024.
(3) Midjourney V6 Alpha
- V6 is the biggest leap from v3 to v4, with the greatest improvement of any version.
- V6 alpha will be released soon. To test it, users must enable V6 manually. V5 remains the default version until V6 becomes the default in January 2024.
- After a week of release, V6 alpha will change dramatically without warning. Later, it will become a more stable beta version with change notifications.
- Many images that did not work in V5 will work in V6.
- The following features are supported in V6: creating text, creating multiple subjects, better control of the composition, image prompting, better anime creation, better remixes, and parameters (chaos, weird, raw, aspect ratio).
- V6 can produce better images of dolphins, whales, dinosaurs, and spaceships. However, for unknown reasons, a centaur is still better in Niji.
- There will be two new upscalers: Subtle and creative upscalers. The upscalers are used for greater control rather than increasing resolution.
- A creative upscaler is more likely to solve an image’s problem. Users can rerun it because the results vary each time.
- After launch, a slider may be added to the upscalers and variations for controllability.
- Some V5 features will be temporarily unavailable during the upgrade.
- There will be no pan or inpainting in V6 at launch.
- The documentation may be unable to keep up with the release of the V6 alpha.
- Because the version still needs to be optimized, V6 will consume more GPU hours.
- V6 is more precise in responding to the prompt. The exact cutoff of prompt length is unknown.
(4) Other updates unrelated to Office Hours
(4.1) Contrary to the claims of AI art haters, AI models do not reproduce the same artwork on which they were trained.
- I want to share a fascinating paper I read in 2023. It’s not the most recent paper (published on 30th January 2023), but it’s a handy reference for when relatives or friends may ask you about your AI work over the holidays.
- The purpose of the paper by Carlini et al., 2023 is to raise awareness about privacy concerns in diffusion models because these models can memorize individual images from training data, which raises concerns about data privacy and the potential misuse of personal information.
- The AI model acts similarly to a human artist who examines images and then learns the patterns that define each subject and style. All of these patterns are then stored in the model’s “memory” (an abstract representation of data or latent space) and used to generate images when users request them.
- According to the findings, AI models can only generate approximate representations of the training images, implying that AI may not be able to reproduce the artwork on which it was trained.
- When researchers attempted to replicate 350,000 images from Stable Diffusion’s training data, they only succeeded with 109 of them, a meager success rate of 0.03%.
- The paper contributes to the ongoing debate on AI and copyright by highlighting AI models’ inability to reproduce the same artwork on which they were trained.
- Reference: Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D. and Wallace, E., 2023. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23) (pp. 5253–5270).
Geeky Animals’ most recent stories:
Tips to achieving style consistency across prompts when using Style Codes.
Previous Quick Midjourney Update and a technique to add an anime character into a photographic image.
Other stories
Latest news & updates on Midjourney:
Simple and easy-to-follow guides for Midjourney Beginners
I hope you like this story!
Please give me a clap or leave a positive comment to let Medium’s algorithm know that this story is helpful and has contributed to its platform’s success.
Please share the Friend Link with your friends, family, and fans to give them free access to this story.
Your support inspires me to create higher-quality content in the future!
Never miss a story. Follow me on Medium for updates, stories, and tips about Midjourney.
Sign up here for a free email alert when I post a new story.