Midjourney's editor replaces Omni-Reference with a newer edit model

• Selected-pixel inpainting has joined the editor, and Midjourney’s release post says it changes only pixels selected by the user.
• The Edit Model replaces Omni Reference, and Midjourney’s reply says it should improve on the old workflow.
• A revised --tile ships in V8.1 and V8.2, while the real-time models remain an early alpha preview, according to the release post.
• A Live Preview recording posted by ai_artworkgen shows four variations forming while a prompt is still being typed.

Midjourney’s Edit Model documentation says one generation can use up to four reference images. Midjourney puts a red hat on a character and carries that character through new settings in its character-edit demo. Four variations form during prompt entry in ai_artworkgen’s Preview recording, before one is selected for upscaling.

Midjourney says its updated inpainting changes only the selected pixels in the release post.

ai_artworkgen ran a three-step edit sequence in their post:

• “change the outfit to camo printed cotton”
• “adjust the lighting to very dark”
• “adjust the lighting to subtle purple neon”

Midjourney says the Edit Model replaces Omni Reference in its reply about the change.

The Edit Model docs define three jobs:

• Change an existing image with written instructions.
• Generate a new image from up to four reference images.
• Edit a selected region or expand an image beyond its original canvas.

The V8.2 Version notes add Character Reference and Retexture to the systems the Edit Model replaces. Midjourney called the interaction a “generic language interface” in its support reply, where it said image2, image two, and perhaps @ image2 should all work.

Midjourney says the new --tile revision is available in both V8.1 and V8.2 in the release post. Its --tile documentation describes a parameter that matches image edges so repeated copies read as one continuous pattern.

Midjourney calls the real-time models an early look on its alpha site in the update.

ai_artworkgen asked for prompt tests against styles they had liked or discovered, plus an aspect-ratio minimum, in a reply on the Preview.

In response to reports that the replacement may perform worse, Midjourney asked users to post examples for its team in a feedback reply.

Midjourney characterized the model as working in a similar way but more generally, and said it was unsure why it would be worse in any situation.