Midjourney Opens V8.2 Image-Editing Tests With Up to Four References

Midjourney has begun broader testing of its first V8.2 image-editing model, giving users one place to alter an image with written instructions, use as many as four image references, and change or extend portions of a canvas. The release shifts several creative controls into an edit-focused workflow, but Midjourney is also asking users to surface cases where the model does not behave as expected.

One editing model, several inputs

The V8.2 model accepts natural-language editing instructions. It also generates images from other images, allowing up to four references at once. Midjourney says this replaces its omni-reference feature in the new editing workflow.

The model includes inpainting, which changes selected areas of an image, and outpainting, which expands its canvas. It also supports personalization, moodboards and srefs. Midjourney cautions that moodboards and srefs may need extra prompt direction in this version.

Ways to access the model

Users can reach the model by attaching images to the prompt bar, selecting Edit from an image lightbox, or using the Edit tab to upload an image. It is available on Midjourney’s main website and alpha site, as well as through Discord.

--edit url...

The multiple access routes place editing alongside Midjourney’s existing prompt and image-viewing surfaces rather than confining it to one new destination. The company has updated the interfaces on both sites, and says the alpha version will change rapidly as it experiments with how the model should be used.

Testing is part of the release

Midjourney is presenting the rollout as a community test, not a settled interface or behavior set. It expects edge cases and directs users to flag poor results through a sad-face control in the lightbox or by sharing the affected image’s job ID or URL while it is open.

The company is also requesting feedback on the user interface and directing ideas and edge cases to its ideas-and-features channel. It says fast-follow updates are planned after the initial test begins, leaving the pace and substance of those changes as the next practical question for users trying the model.