Midjourney adds self-correcting "thinking" reruns
Midjourney now lets alpha users rerun V8.2 images in a "thinking" mode that checks the output against the prompt and generates again, a small move toward self-critique in image generation.
Midjourney announced a "thinking mode" for jobs made with V8.2 in its weekly changelog for alpha.midjourney.com. The mode works on both standard and edit jobs. The company says the system looks at the generated image, finds where it missed the prompt, and generates again. It names four kinds of miss: wrong objects, layout, anatomy and text.
Midjourney calls it "an early version" and is asking for feedback. The changelog gives no benchmarks, no figures on how often a rerun improves on the first image, no added cost or generation time, and no details about the model that does the checking.
The changelog's main item is shared folders. Owners can invite people by username as Collaborators, who can generate straight into the folder, or as Viewers, who can only browse. Folders can also be shared by link with three access levels: private to invited members, viewable by anyone with the link, or open for any Midjourney member to contribute. Images stay in each creator's own history. Midjourney warns that sharing doesn't change image privacy. Non-stealth images made inside a private folder can still appear on Explore and on profiles.
Other changes were mostly fixes and interface work. Stalled edit submissions now fail cleanly within two minutes. Large moodboard uploads now show a progress bar. Draft mode on V8.1 and V8.2 always produces a 24-image grid, and the interface now says so. Suspended accounts get an Appeal button. Midjourney says persistent edit history and more style tools are coming next.