Midjourney, OpenAI and Google Are Racing to Fix AI Image Editing

AI image generation is already good enough to impress almost anyone. You can make a nice fashion campaign, a product cool advert, a cinematic portrait or a poster in less time than it takes to explain the idea to a designer.

Then you try to edit it.

You ask the model to make the jacket black, and it changes the face, moves the background, replaces the watch and gives the whole image slightly different lighting. The result may still look beautiful, but it is no longer the picture you were working on.

That is where the real AI image race is happening now. Midjourney, OpenAI, Google, Black Forest Labs and Adobe are no longer competing only over who can generate the prettiest first attempt. They are trying to build an editor that can understand a simple instruction such as “change this one thing and leave everything else alone.”

It sounds basic. None of them can do it perfectly yet.

Quick Answer

There is no officially announced GPT Image 3, Midjourney V9 or Nano Banana 3 release date that we could verify. Midjourney has already moved through V8 and V8.1 alpha releases, OpenAI currently lists GPT Image 2 as its leading image-generation and editing model, Google has Nano Banana Pro, and Black Forest Labs already offers FLUX.2 and FLUX.1 Kontext. (Wikipedia)

The next meaningful improvement will not be another model producing a slightly prettier woman in cinematic lighting. People want accurate local editing, stronger identity preservation, several reference images and characters that remain recognisable after multiple rounds of changes.

Our personal favourite is still Nano Banana Pro, mainly because its photorealism and ability to keep an image feeling natural are difficult to beat. GPT Image 2 is also working very nicely, particularly when you need to make quick graphic-design changes without opening Photoshop and moving everything manually.

Google Currently Feels Closest to a Proper Photo Editor

Nano Banana Pro is the model we currently trust most when the final image needs to look like a photograph rather than an obvious AI composition.

It is particularly good at taking existing people, objects and backgrounds and fitting them together without making the final image feel completely rebuilt. Google says Nano Banana Pro can combine up to 14 input images, maintain the resemblance of up to five people, perform local edits and output images at 2K or 4K resolution. It can also adjust focus, lighting, camera angles and colour grading through normal instructions. (blog.google)

That does not mean every edit lands perfectly. Faces can still drift, repeated changes can gradually damage the original image, and the model occasionally decides that your simple request needs a full creative reinterpretation.

Still, when it works, Nano Banana Pro gets closer than most tools to the feeling of telling a human editor what you want. You can provide several references, explain the scene and continue making changes without beginning from zero every time.

The photorealism is the main reason it remains our favourite. It often understands how people, light, skin and camera depth should fit together without producing that polished plastic quality many AI models still fall into.

Google is not waiting around either. The company has treated image editing as a central Gemini feature rather than an extra button beside image generation, and Nano Banana Pro is already available across products including Gemini, Google AI Studio and parts of Workspace. (blog.google)

GPT Image 2 Is Excellent When You Need to Move Fast

GPT Image 2 feels slightly different.

Nano Banana Pro is usually our first choice for realism, while GPT Image 2 often becomes more useful when the task sits somewhere between image editing and graphic design. OpenAI describes it as its current state-of-the-art model for fast, high-quality image generation and editing, with support for both text and image inputs. (OpenAI Developers)

The biggest advantage is the conversational workflow. You can upload a design, explain what needs to move, ask for another format, change the headline, replace one element and keep working through the same chat.

For quick social graphics, thumbnails, article images, mockups and layout ideas, that is extremely useful. You do not always need pixel-perfect Photoshop control. Sometimes you need to move the product to the left, create more empty space for a title and produce a square version without spending forty minutes rebuilding the composition.

GPT Image 2 is good at that kind of work because the model understands the request in context. You can explain why the previous result was wrong rather than writing a fresh prompt and praying that the next generation somehow remembers what you liked about the first one.

Its weakness is similar to every other conversational editor: the longer you continue, the more likely it becomes that something starts drifting. A face changes slightly, small objects disappear or the design becomes cleaner in places where you did not ask it to become cleaner.

The next OpenAI image model does not need a dramatic new name. It needs stronger identity locking and the ability to survive a long editing conversation without slowly replacing the original.

Midjourney Still Has the Best Eye in the Room

Midjourney remains the model people open when they want an image to look expensive.

Its greatest strength has always been taste. Give it a rough idea and it often produces better composition, mood, colour and lighting than you knew how to request. That quality made Midjourney famous, but it is also why editing can become frustrating.

Sometimes you do not want Midjourney’s opinion.

You already like the image. You want the same person, same framing and same atmosphere, only with a different shirt or another object added to the table. Midjourney has editing, region variation, reference and remix tools, but it still has a habit of treating a small change as permission to improve the whole picture according to its own taste.

V8 and V8.1 have already arrived in alpha form, so the old story about everyone waiting for V8 is no longer current. The more interesting question is whether Midjourney’s next updates can reduce the famous Midjourney randomness without removing the visual character that makes people use it in the first place. (Wikipedia)

Midjourney does not need to become Photoshop. It does need to become more obedient.

FLUX Is Building the Serious Editing Engine

Black Forest Labs does not always receive the same consumer attention as OpenAI or Google, but FLUX may become just as important behind the scenes.

The company currently offers FLUX.2 Max, FLUX.2, the faster FLUX.2 Klein and FLUX.1 Kontext. Black Forest Labs describes FLUX.2 and Kontext as in-context generation and editing systems that combine text and image inputs for coherent edits, while Klein is designed to generate and edit in under a second. (Black Forest Labs)

That speed and flexibility matter because FLUX is likely to appear inside other products rather than forcing everyone into one specific app. Developers can use it for image tools, ecommerce, advertising, character workflows and more controlled creative systems.

FLUX.1 Kontext was built specifically around the problems everyone is now trying to solve: local editing, style references, character references and consistency through iterative work. Its research paper reports stronger preservation of objects and characters across multiple turns compared with earlier editing systems. (arXiv)

The model still has the same long-term weakness as its competitors. Every additional edit creates another chance for the image to degrade or drift, but FLUX is one of the clearest signs that AI editing is becoming its own category rather than a secondary feature of text-to-image generation.

Adobe May End Up Winning Without Building the Favourite Model

Adobe does not need everyone to agree that Firefly is the best image generator.

It already owns Photoshop.

That position allows Adobe to take models from Google and Black Forest Labs and place them inside a proper editing workflow with selections, masks, layers, scaling, compositing and manual controls. Photoshop currently supports partner models including Nano Banana Pro and FLUX inside Generative Fill, alongside Adobe’s own Firefly models. (Adobe)

This may be more important to professional creators than another standalone model winning a leaderboard.

A designer does not only need a nice final JPEG. The file has to remain editable. Text needs to stay movable, objects need separate layers, masks need to be adjustable and another person may need to open the project next week without discovering one flattened AI image and no way to change anything.

Adobe is adding AI to the place where that work already happens. Photoshop can now use Generative Fill, reference images, object rotation, Harmonize, background generation, expansion, upscaling and automatic layer cleanup while still leaving the user inside Photoshop. (Adobe)

The future may not be Adobe defeating Google or OpenAI with one superior model. It may be Adobe allowing you to choose the model you want, then giving you the proper tools to repair whatever it gets wrong.

What Everyone Is Actually Waiting For

The AI companies already know how to make beautiful images. The unfinished problems are much less glamorous.

The first is real local editing. When you select a shoe, the model should change the shoe without redesigning the leg, floor and person wearing it.

The second is identity locking. One person should remain the same person across different poses, clothing, lighting and ten rounds of edits, rather than slowly turning into a similar-looking cousin.

The third is multi-reference control. Users should be able to provide a face, outfit, product, location and visual style, then tell the model exactly which qualities to take from each image.

The fourth is stable multi-turn editing. Changing the background in step seven should not erase the necklace added in step three or damage the face established in the original photograph.

The final piece is editable structure. For serious design work, the result should eventually arrive with useful layers, objects, text and masks rather than one beautiful but completely flattened image.

Those improvements would matter more than another jump in raw generation quality because the current models are already capable of creating something impressive. The frustration comes when you need to use it for actual work.

The TGK Take

Midjourney, OpenAI and Google are racing toward the same goal from different directions.

Midjourney still has the best instinct for style and atmosphere, Google currently gives us the most convincing photorealistic editing, and GPT Image 2 is excellent when you want to make quick graphic-design changes through a normal conversation. FLUX is building the underlying editing technology, while Adobe is placing all of it inside the software professionals already use.

There is no confirmed GPT Image 3, Midjourney V9 or Nano Banana 3 waiting behind a curtain with a release date. The current tools are already strong enough that the next improvement needs to be about control rather than another round of prettier demo images.

The company that finally understands “change this and do not touch anything else” will have something much more useful than the best AI image generator.

It will have an actual editor.