Forum

Three Days in July: Meta’s Muse Image and the Cost of Default Enrollment

A row of pale blue wooden benches lined up along a tall green hedge beside a gravel path

On 7 July, Meta introduced Muse Image and previewed Muse Video, the first image and video models built entirely in-house by its Superintelligence Labs. Muse Image arrived across the Meta AI app, on meta.ai, in Instagram Stories in the United States and on WhatsApp in some countries.

One feature turned the launch into a lesson in consent. On Instagram, users could mention a public account in a prompt, and the system would pull that person’s photos as visual references for a new image. Adult public accounts were included by default, and the people whose photos were used were not notified. The reaction was fast. The performers’ union SAG-AFTRA called anything short of a clear, conspicuous opt-in unacceptable and “an utter miscalculation of public sentiment”. On 10 July, Meta removed the feature, saying it had missed the mark. Muse Image itself remains available.

A day after Meta’s launch, on 8 July, ByteDance released Seedream 5.0 Pro, an image model built around a very different idea of what designers need: it can split a generated image into separate, editable layers, edit specific regions precisely and render text in many languages.

The consent problem, again

The Muse episode repeats a pattern this column has followed for a year. A powerful capability ships with permission assumed, public reaction forces a retreat, and the company presents the change as a response to feedback. The specific twist this time was scale and intimacy. Public Instagram accounts belong to ordinary people, not only to celebrities, and “public” has never meant “available to be remixed into new images by strangers”.

For anyone designing products that touch personal images, the lesson should by now be settled. Opt-in for anything involving identifiable people is not a nice extra; it is the minimum that users, unions and regulators will accept. Notification when your image is used is part of the same principle.

Layers instead of pictures

Seedream 5.0 Pro deserves attention for a different reason. Its layer separation, breaking a render into independent transparent elements such as a line of text, a character, an object or the background, turns a generated picture into something closer to a design file.

That matters for professional work. Designers rarely need a flat image; they need assets they can arrange, adjust and reuse within a system. Layered output fits the way design actually happens, and it makes human editing a natural next step rather than a workaround.

Two directions for image AI

The two launches point in opposite directions. One treats the social graph as raw material, pulling in other people’s photos to make generation feel personal. The other treats the designer’s workflow as the destination, producing material that people can take apart and rebuild. The second direction is more respectful of both creators and subjects, and probably more useful.

Practical steps for creatives

  • Check the privacy and AI settings of any public accounts you use for your portfolio, and review them after major product launches.
  • Do not use tools that pull other people’s images without their clear permission, however convenient.
  • Prefer tools that output editable, layered assets when the work needs to live inside a design system.
  • Record which parts of a composition came from a model and which were edited or created by hand.
  • Talk to clients about it. Many will not know these settings exist, and a short conversation now can prevent an awkward discovery later.

The award perspective

The International AI Award Rules page asks entrants to address ethical considerations and to own, or have permission for, everything they submit. Work built on strangers’ photos drawn in without consent would fail those tests, however striking. Work that uses layered generation as a starting point for careful human design is the kind of practice the field should encourage.

The takeaway

In three days, Meta demonstrated again that consent cannot be retrofitted. In the same week, a competitor showed that generative tools can be built around the way designers work. The future of image AI will be decided less by model quality than by which of those two instincts wins.