THE SHORT VERSION
Midjourney's first V8.2 edit model is open to every account. The craft shift: when editing is a prompt, the working file stops being the output and starts being the input.
• Opened August 27: instruction edits, inpainting and outpaint, and up to four image references in one generation, replacing omni-reference.
• Access is in the web prompt bar, the lightbox edit control, the edit tab, or the edit parameter on Discord.
• The habit that transfers from older tools: one instruction per pass, verify, then the next instruction.
• Caveat: this is Midjourney's own announcement of a rolling test, and the post itself asks for help with edge cases. Vendor claims, vendor source.
ONE · WHAT SHIPPED
Midjourney posted on August 27 that its first V8.2 image edit model is open to everyone for testing. The post lists four capabilities: editing an image by instruction, generating from other images with up to four references at once, inpainting specific areas or expanding the canvas, and combining the model with personalization, moodboards, and style references. The moodboard and style-reference paths carry an honest note that they may want some extra prompt direction in this version.
Four routes in: attach an image under the prompt bar, use the edit control in the image lightbox, open the edit tab and upload, or run the edit parameter with a link on Discord. The company says both the main site and its alpha interface will change rapidly, and it is asking the community to report edge cases with the reaction control in the lightbox.
Read the status words exactly. Opened to everyone for testing is a rolling release, and the same post promises fast-follow updates. The capability is real and reachable today; the polish is not finished, and the maker says so.
GO TO THE SOURCE
The August 27 post is on Midjourney's own updates page, here. It is the maker announcing its own model, so treat the capability list as vendor claims and the interface notes as the part most likely to move.
TWO · THE EDIT PROMPT IS A DIFFERENT SHAPE
A generation prompt describes a whole frame. An edit prompt describes a difference. "A ceramic mug on stone, morning light" is a generation. "Change the surface to weathered oak, keep everything else" is an edit, and its subject is the change, not the scene. The scene is already in the file you attached.
The discipline that transfers from every editing tool is one instruction per pass. Ask the model to move the light and swap the surface in one instruction and you cannot tell which instruction did the damage when the result drifts. Edit, verify against what you wanted, then issue the next instruction.
The subtraction habit applies with one change. In a generation, you name what must not appear. In an edit, the instruction itself is a subtraction: remove the reflection, drop the background figure. The negative list becomes the work order.
THREE · FOUR REFERENCES REPLACE OMNI
The edit model takes up to four image references in a single generation and, per the post, replaces omni-reference. That moves character and series work one step: the reference discipline from earlier issues now spans four inputs, which means the anchors have to be named, not implied.
The working split I would test first: one reference for the subject, one for the surface, one for the light, one for the palette. Four slots is a budget, and a budget forces the question of which reference is doing which job. If two references carry the same job, one is a duplicate and the fourth job is going without.
FOUR · WHERE AN EDIT BEATS A REROLL
A reroll is still the right tool when the composition itself is off. But for the failures that used to cost a full regeneration, an off surface, a stray reflection, a background that grew a fence, an instruction edit keeps the frame you already liked and changes one thing in it. The revision issue's rule holds: change one thing, keep the rest, verify.
The honest limit: this is day one of a rolling test, and the post says so. Expect edge cases, expect the interfaces to move, and expect the instruction-following to sharpen over the fast-follows. Test it on the failure case you would otherwise have re-rolled, and keep the pair of frames so you can see what the edit actually preserved.
PROMPT OF THE DAY
Reach for this shape on any image that is one detail away from usable. Attach the frame, then issue the change as a single instruction with the keep-clause attached.
Change the surface under the mug to weathered oak, visible planks, same daylight direction. Keep the mug, the camera angle, and the background exactly as they are.
The load-bearing line is keep the mug, the camera angle, and the background exactly as they are. An edit prompt without a keep-clause invites the model to improve everything, and the frame you liked is gone.
START HERE
Take the last image you re-rolled more than twice. Attach it in the edit model instead, and write the one instruction that describes the difference you wanted. Compare the pair. If the edit model keeps what the reroll kept losing, that frame is your proof.
•••
I am building an instruction edit kit: the edit shapes that hold across product, portrait, and landscape work, each with its keep-clause and the failure it prevents, tested against the first V8.2 edit model.
Want it when it ships? Reply with send me the instruction edit kit and I will get it to you.
A QUESTION FOR YOU
Which re-roll in your last session was really an edit in disguise?
A good answer names the frame, the one detail that was off, and whether a keep-clause would have saved the reroll.
If this was useful, forward it to a creator who rerolls five times before editing once.
Until next time,
Luxe Prompting
Luxe Prompting
AI Systems and Prompt Craft for Creators
| 
