How to Edit an Image in Midjourney: The Edit Model

Editing in Midjourney changed with V8. The separate tools many guides still describe, Character Reference, Omni Reference and Retexture, have been folded into one Edit model that works with V8.1 and V8.2. It does three jobs: it changes an existing image from written instructions, it builds new images from up to four reference images, and it powers the Editor's inpainting and outpainting. This guide covers each of them, with the limits worth knowing before you spend GPU time.

Three ways into the Edit model

1. Quick Edit. Open any image from your Create or Organize page and look under Creation Actions. Quick Edit puts the image into the Imagine bar so you can type an instruction.
2. Attach to prompt. Click the image icon in the Imagine bar and drag up to four images into the Attach to prompt section. Lock them with the padlock if you want them to stay for several prompts.
3. The Editor. Choose Open Editor from Creation Actions, or open the Edit tab in the website menu to work on uploaded images and use the advanced tools.

The Edit model also works alongside image prompts, style references, moodboards, personalisation and --hd. Two exceptions at the time of writing: it is not compatible with --tile, and its results cannot be used with Remix.

Editing with written instructions

Classic Midjourney prompts describe the finished picture and avoid commands. The Edit model is different: Midjourney's documentation says you can give it direct instructions, much like Conversational mode. Both styles work, so try each.

• Instruction: "make this a winter scene, snow on the roofs, overcast light".
• Description: "the same street in winter, snow on the roofs, overcast light".

Useful jobs for instructions include changing the time of day or weather, swapping a background, turning an illustration into a photograph ("make this girl into a real character" is Midjourney's own example), and showing another viewpoint: "let's see this image from behind".

Two settings matter. The Edit model tries to match the aspect ratio of the first image you attach and ignores your default ratio, so add --ar to the prompt if you want a different shape. And adding --raw reduces Midjourney's own flourish so the instruction is followed more literally.

Reference images and consistent characters

The Edit model accepts up to four reference images and replaces both Omni Reference and Character Reference for V8. That makes it the tool for putting the same person, product or creature into a new scene, or for combining several elements, as in Midjourney's example of a silver knight riding a horse of fire while holding a sword of flowers.

1. Attach a clear image of each character or object, ideally on a plain background.
2. Describe the new scene in the prompt, naming the elements simply ("the knight", "the horse").
3. If features of two characters get mixed up, Midjourney suggests combining them into a single reference image.
4. Keep the references and version fixed across a series so the character stays stable.

Our guide to consistent characters covers the wider workflow; the Edit model is now the V8 way to do it.

Restyling an image

Retexture used to be a separate Editor tab. In V8.x it is part of the Edit model: attach the image and describe the new medium, optionally with a style reference, moodboard or personalisation profile. Midjourney's example is "hand drawn pencil sketch of this image" with an --sref code, and its advice is to describe the style in words as well, which helps the reference take hold. The old Retexture tab is still available if you set your default version to V7 or earlier.

The Editor: brushes, Smart Select and layers

The Editor is where you change one part of an image (inpainting) or extend the canvas (outpainting). Opened from your gallery it starts as a light version; the Edit tab has the full set.

With Smart Select you click to set include points, add exclude points where the mask spreads too far, then press Erase Selection or Erase Background before submitting. With layers, only the transparent, erased areas are regenerated; visible parts stay unchanged. To regenerate the whole image in one consistent style, use a restyle instruction instead.

Edits on HD images return SD results, so edit first and upscale afterwards. Midjourney lists an Edit model SD prompt at about 1 GPU minute and an HD prompt at about 2.3 minutes, more than a normal prompt.

Privacy of edited images

This catches people out. In V8.x, creations made in the Editor with the Edit model appear on your Create and Organize pages and are visible on midjourney.com to other members unless Stealth mode is on. That applies to your uploaded images too. In V7 and earlier, edits made with uploaded images stayed visible only to you. If you edit client material, switch Stealth on first (Pro and Mega plans), or keep that work out of Midjourney.

Try it

make this a rainy evening, wet pavement reflecting the shop lights, keep the people and the composition the same --raw

Use it on a daylight street scene with Quick Edit. If the people change too much, open the Editor instead, erase only the sky and pavement, and give the same instruction, so the rest of the frame is protected.