Image editing models: which engines can modify without redoing everything
Image editing starts from an existing visual and transforms it under constraints: keep the character, change the light. The methods (image to image, instruction, mask, references), the families that can edit, and why control is what is at stake.
An image editing model starts from an existing visual and transforms it while respecting constraints: keep the character, the product, the framing, and change the light, the clothing, the setting, a material. Three methods coexist (image to image, editing guided by a natural-language instruction, retouching by area with a mask), often combined with references. In 2026, FLUX.2, Qwen-Image-Edit, Gemini Image and GPT Image can edit, each in its own way; at Axiris, what is at stake is not the transformation but the control: changing what has to change without destroying what is validated.
Modifying an image with AI
Image editing starts from an existing visual and asks the model to transform it. The intention is no longer "create an image" but "transform this image while respecting these constraints". For a graphic designer, that difference is fundamental: this is where professional work begins.
This page answers one question: which engines can edit, and how. The gesture itself (starting from an image, retouching an area) is described in Transforming an existing image and Retouching by area.
Keep and transform
An editing job asks you to keep the character, the composition, a product, the framing, a face, elements of the setting, while modifying the light, a garment, the environment, a material, a colour, an object, the style.
What matters is not the list of what a model can change. It is what it can leave unchanged.
Three methods
Image to image uses the source image as the base for a new generation: the model departs from it to a greater or lesser degree, according to a setting.
Editing by instruction directly interprets an instruction in natural language ("remove the background", "change the light", "take out the passers-by") and preserves the subject, the scene, the pose.
Retouching by area uses a mask to delimit where the engine is allowed to work; the rest of the image is kept.
These methods combine with references: one image for the subject, one for the style, one for the setting. How you give each one a role: Workflow with visual references.
The families that can edit
In 2026, several families offer editing, each with its own approach: FLUX.2 (Black Forest Labs: text-guided editing, multiple references; the klein variant is our engine for retouching by area), Qwen-Image-Edit (the editing branch of Qwen-Image), Gemini Image (natural-language, multimodal editing) and GPT Image (text and images in context). Ideogram 4.5, announced in October 2026, goes the same way: mask, up to four references, layout by structure. See our reading.
This list does not rank. Each engine edits differently; what settles it is a test under equal conditions: Comparing AI image models.
Control is what is at stake
The more professional the work becomes, the more the question becomes: what can we change without destroying what already works?
That is the Axiris method: the image is built by areas, each area is a box with its mask, and what is validated is frozen as a layer and no longer moves. See Building the image by areas. In Nova, editing by instruction and compositing several references have been in service since 1 July 2026 (the news item), and reworking one element without touching the rest since 13 July 2026 (We no longer generate an image. We build it.).
Contact request: a first conversation, free and with no commitment. Every practice is different; it is by listening to you that we propose what suits you.