Midjourney's October had two very different faces: a steady drip of alpha releases, including a thinking mode that re-checks its own work, shareable folders, and sharper edit tools, plus a copyright report that names Midjourney v8.1 the worst offender among the major image and video models.
Three threads mattered most this month: the VN.ai copyright report, the new Thinking Mode test, and an alpha changelog that quietly shipped real collaboration tools.
A Report Names Midjourney Its Worst Offender
VN.ai, a startup that stress-tests generative AI tools for copyright infringement, published its first quarterly report in mid-October. It ran 15 popular models through prompts designed to surface studio IP, scoring them on how often they reproduced protected characters and how often those characters appeared in "brand safety violations" - protected figures placed in harmful contexts such as violence or drug use.
Midjourney v8.1 finished at the bottom. According to the report, which was shared with Deadline, the model produced images that violated brand safety measures once in every three text prompts, or 37% of the time. One cited example was a v8.1 hallucination that generated The Joker smoking cannabis while wearing a dress the report described as something that could have been ripped from Frozen's Elsa wardrobe.
Midjourney is not alone, and that is part of the story. Every model in the test produced at least one copyright-protected character, including prompts where the character's name never appeared.
- MiniMax H3 generated videos that broke studio brand safety standards 55% of the time when an image was part of the prompt.
- Google's Nano Banana produced at least one protected character.
- ByteDance's Seedance 2.5 still produced protected characters, though the report credits it with improving significantly after Hollywood studios sent cease-and-desist letters.
The legal backdrop has not moved. Disney and NBCUniversal sued Midjourney more than 18 months ago, accusing it of being a "bottomless pit of plagiarism," and the case remains ongoing. Warner Bros Discovery has filed its own suit. Midjourney has argued that training its models on copyrighted images is permitted under fair use, and it did not respond to requests for comment on the report.
VN.ai founder Nelson Chu offered a blunt explanation for why models keep drifting toward recognizable characters. "It's good for business," he told Deadline. "It's great marketing. It works. It gets people to use your model and pay for it." He also predicted the pattern will fade, arguing that as more models set a precedent for being a good actor, the rest will improve.
The practical takeaway for anyone piping generated images into commercial work is unchanged, only sharper. Treat every output as unverified until you have looked at it, and never prompt by franchise or character name, because the report suggests a clean prompt is a floor rather than a guarantee.
Thinking Mode Asks the Model to Check Its Own Work
On October 8, Midjourney opened a "Thinking Mode" test on its alpha site. The mechanics are simple: open a job in the lightbox on alpha.midjourney.com and click "Rerun (Thinking)" in the lower right corner.
The changelog describes what actually happens. Midjourney looks at your existing image, finds the places where it missed your prompt - wrong objects, broken layout, bad anatomy, mangled text - and generates again with those misses in mind. It works on jobs made with V8.2, both standard and edit.
In Midjourney's own tests, the rerun helps with prompt accuracy, typography, and coherence. That trio is telling: typography and text rendering have long been the weak spot of image models, and prompt adherence is what separates a usable generation from a lucky one.
It is an early version, and the company is explicit that it wants feedback, especially examples of where thinking helps and where it does not. Longer term, Midjourney says it may ship this broadly as a thinking mode, or as a way to "add more thinking" after a job already exists. Either way, the framing is familiar: the image model gets a second pass at its own output before you accept it.
Folders, Previews, and Cleaner Edits
The quieter half of the month lived in the alpha changelogs. The October 7 entry led with shared folders, the first real collaboration feature on the platform. You can invite people by username as Collaborators or Viewers, share by link, and choose whether the folder stays private, is view-only for anyone with the link, or is open to any Midjourney member. Collaborators generate straight into the folder and build on each other's work, while the owner can delete a shared folder for everyone. Images still stay in each creator's own history.
One caveat is worth repeating: sharing does not change image privacy. Non-stealth images can still surface on Explore and on profiles even when they were made inside a private folder.
The rest of the month's alpha work was unglamorous and genuinely useful:
- Improved inpainting and outpainting from the edit model, where targeted edits now change only the pixels you selected, so repeated passes stop degrading the image.
- Live Style Previews in the Styles sidebar, showing what your latest prompt looks like across liked and featured styles, plus an experiment with faster models in the interface.
- Seamless --tile blending for V8.1 and V8.2, fixing the visible seams between tiled images.
- Per-folder default parameters, style-reference pills grouped into a single pill with thumbnails, and a fix for a crash that hit users opening an image deep inside a 100,000-item history.
Underneath all of it sits V8.2, the default image model since July 24, 2026, focused on aesthetics, image quality, personalization profiles, and native 2K output.
The October 1 changelog signed off with an ambitious phrase, describing better ways to explore, find, and create styles as moving "towards what David calls the aesthetic singularity." It is a grand target. In the same month, a copyright report put a hard number on how much these models lean on other people's characters to get there.
Comments