Midjourney brings text-guided editing and multiple references to V8.2

The new V8.2 edit model modifies images through instructions, combines up to four references, and supports inpainting and outpainting.

Midjourney has opened its first V8.2 edit model for community testing. The release brings instruction-based editing, regional replacement, canvas expansion, and generation from multiple images into a single environment.

The model accepts up to four references at once. Midjourney presents this feature as the successor to Omni Reference, which was limited to one image in V7. References can represent characters, objects, clothing, environments, or other elements that users want to incorporate into a new composition.

Text-guided editing starts with an imported image and a written description of the requested changes. Users can modify the entire scene or select a specific area with a brush to replace an individual detail. Outpainting extends the original frame by generating new content around the existing image.

The model also works with account personalization, moodboards, and `--sref` style references. Midjourney recommends adding further written guidance when using these options, suggesting that their interpretation remains inconsistent in this early version.

There are several ways to access the model. Users can drop an image into the prompt bar through “Attach to prompt,” select “Edit” from an image in the gallery, or upload an external image through the website’s dedicated Edit tab. On Discord, the `--edit` command followed by an image URL launches the same process.

The interfaces on midjourney.com and alpha.midjourney.com have been updated to support the new workflow. The Alpha website will continue serving as an experimental space for new approaches to selection, composition, and reference control.

Midjourney describes the launch as a community test rather than a stable release. Users are being asked to flag failed generations from the gallery and share the corresponding job IDs. The company is looking both for cases where the model behaves unexpectedly and for feedback on the interface itself.

Less than 24 hours after opening the test, Midjourney released its first image-quality update. The company invited anyone who encountered problems during the initial rollout to try their requests again. It did not disclose the technical nature of the changes.

The quick revision shows both the appeal of the new system and its unfinished state. Support for four references moves Midjourney further toward controlled image composition, while inpainting and outpainting already existed in its editor through separate tools. The more significant change is their consolidation within a V8.2 model that can receive several source images and a shared instruction.

Midjourney has not yet detailed the computing cost of these operations, their compatibility with every V8.2 parameter, or any differences between results produced on the website and through Discord. Additional updates are expected in the near term.