Image-to-image AI: how it actually works

The phrase “image-to-image AI” covers more useful ground than almost any other term in generative AI, and most of the articles ranking for it explain less than the average YouTube tutorial. This is the long version: what it actually is, how it works underneath the marketing copy, which tools are worth opening today, and how it changes the practical jobs people do in property, design and trades.

If you have a picture and you want a better picture, or a different version of it, or the same picture with one thing changed, this is the technology you are reaching for. Text-to-image is the cousin that makes things from nothing. Image-to-image is the one that keeps the room and changes the sofa.

What image-to-image AI actually is

Image-to-image AI is a generative system that takes two inputs, an existing image and an instruction, and returns a new image that retains the structure and composition of the input while applying the requested change. The input image is the anchor. The instruction, usually a written prompt and sometimes a mask or a reference image alongside it, is the change.

The most useful way to think about it is by contrast with text-to-image. Text-to-image gives the model a blank canvas and asks it to invent: a sunset over Bondi, a Hamptons kitchen, a person eating lunch. The model decides the angle, the lighting, the people, the layout, everything. Image-to-image gives the model your photo and asks it to leave most of it alone. Same Bondi photo, but make the sky pink. Same kitchen, but change the cabinetry to white shaker. Same person, but in a different outfit.

That single difference, who chose the composition, is why image-to-image is the workhorse model for any commercial use where the result has to correspond to a real subject. A real room, a real product, a real building, a real person, a real listing. Text-to-image invents. Image-to-image preserves and modifies. Most of the commercial value in generative imagery for property, trades and design sits on the preserving side.

The technical term sometimes used for the underlying capability is img2img (the long-running shorthand in Stable Diffusion communities) or image editing, which is the term Google, OpenAI and Adobe have all standardised on in 2026. They all mean the same thing. The 2026 frontier models call it editing because that is what users were actually doing with it: editing photos they already had, not generating new fantasy scenes.

Before After

How it actually works

This part is short, and reading it is the difference between using these tools well and treating them as a magic black box.

Modern image-to-image models are diffusion models (with a handful of newer flow-matching variants in the latest generation). The mechanism, which sounds backwards the first time you read it, is that the model is trained to remove noise from images. Start with a clean photo, add a controlled amount of random static, then teach a neural network to recover the original from the noisy version. Do that millions of times over millions of images, and the network learns the structure of what photos and paintings and renders look like, well enough to remove noise even from images it has never seen.

To generate a new image from a prompt alone, the model is handed pure noise and asked to denoise it into something matching the description. To edit an existing image, the model is handed your image with a calibrated amount of noise added (not pure noise, but partial noise), and asked to denoise it back to something that both looks like your input and matches your prompt. The amount of noise added is the key parameter. Less noise means the result stays close to your input. More noise means the model has more freedom to reinterpret, and the result can drift further from the original.

That parameter has a name that everyone interacting with these tools should know: denoising strength (also called “strength” or “image weight” depending on the interface). It runs from 0 to 1. At 0, the input image is returned unchanged. At 1, the input is functionally ignored and you are back to text-to-image. Most useful image-to-image work happens between about 0.3 and 0.7, where the original composition is preserved but the model can meaningfully change materials, finishes, furniture or atmosphere. This single dial is the difference between editing a photo and replacing it.

A more recent class of model, including Black Forest Labs’ Flux Kontext family released in 2025 and now used widely in 2026, swaps pure denoising for in-context image generation. The model is given the input image and the prompt together as joint conditioning, and outputs a modified version in a single pass. The user-facing experience is the same, change a photo with a sentence, but the underlying mechanics let it hold character, finish and identity more consistently across edits than older img2img pipelines did.

The frontier models from Google (Nano Banana Pro / Gemini 3 Pro Image) and OpenAI (GPT Image 2) take this further again. They are large multimodal reasoning models that treat image editing as one task among many, which means they understand the request, the existing image and the relationship between them at a higher level than a pure diffusion engine, and they keep the unedited regions of an image pixel-stable while applying the requested change. This is the capability that finally made text inside generated images legible, and that made multi-turn editing (“now make the cushions pink”, “now move the lamp to the left”) feel like a conversation rather than a series of disconnected generations.

What image-to-image AI can actually do

Before After

There are at least eight distinct jobs to be done under the image-to-image label, and conflating them is the main reason people get bad results. The same tool that excels at one is often weaker at another. The capabilities, in roughly the order they get reached for in commercial work:

Restyling and re-decoration. Take a photo of a room and change the entire visual style: turn a tired beige rental into modern coastal, a builder-grade kitchen into Hamptons, a vacant living room into Japandi. Geometry and structure stay locked. Materials, finishes, furniture and palette transform. This is the dominant property use case and the reason the AI-room-design category exists.

Inpainting. Mask a region of an image and regenerate only that region according to a prompt. Used for removing objects (rubbish bins, parked cars, a previous tenant’s belongings), replacing objects (swap the splashback, change the bedhead), or fixing model errors in a previous generation. Inpainting is the most surgical form of image-to-image because the edit is scoped to a defined area, leaving the rest untouched.

Outpainting. The reverse: extend the canvas beyond the original frame, with the model inventing what would plausibly sit just outside the shot. Used for re-cropping a tight photo for a hero slot that needs a wider aspect ratio, or completing a partially obscured building.

Style transfer. Apply the visual style of one image to the content of another. The original neural-style-transfer demos in the late 2010s rendered photographs as Van Gogh paintings. Modern tools use the same idea for practical work: matching a brand palette across a campaign, applying a consistent illustration style to a series of marketing assets, or carrying a single reference look across an entire interior project.

Sketch-to-photo and sketch-to-render. Feed in a rough hand sketch, an elevation drawing, a wireframe CAD export or a basic 3D massing model, and get back a photoreal interpretation. The architectural visualisation tools that have grown up around this capability (mnml.ai, Gendo, Archsynth, Vizcom, PromeAI) sit on top of image-to-image foundations with control nets that lock the line work in place. Useful for concept presentations, fast feasibility imagery and option-generation, less reliable for council-grade photomontages where actual geometric accuracy is required.

Virtual staging. Put furniture into an empty room without owning any furniture. The model preserves the architecture, lighting and finishes of the input photo and adds beds, sofas, dining sets, art and styling that suit the brief. The category has consolidated around a handful of specialised platforms in the past two years (and around generalist multimodal models that now do this competently as a side capability). Pricing has collapsed from $25 to $50 per traditional virtual stage to as low as 50 cents per image on AI platforms.

Background replacement and environment swap. Hold the subject, swap everything else. Real-estate day-to-dusk, e-commerce white-background shots, replacing the bin in the background of a perfect kitchen photo, putting a product on a different surface. Tools have been doing this for a few years; the 2026 quality bar is now close enough to photographic that the cleanup work is in the post, not in the generation.

Restoration and enhancement. Sharpen, upscale, de-blur, colour-correct, fix noisy night-time exposures. Pure photo improvement, no creative change. The tools have largely merged with general image-to-image now that the same model architecture handles both jobs.

The right tool depends on which of these you are doing. Generalist tools like Nano Banana Pro and GPT Image 2 cover almost all of them at usable quality. Specialist tools beat them in their niche: virtual-staging platforms for batch property work, sketch-to-render platforms for architects, Flux Kontext for character and identity consistency.

The tools that lead the category in 2026

The image-to-image tool landscape has consolidated dramatically in the past eighteen months. The list below is the set that matters today, with the actual job each one is best at. Skip past it if you only want the concepts.

Generalist frontier models

Google Nano Banana Pro (Gemini 3 Pro Image) is the current leader for edit fidelity and prompt-following on consumer-facing tasks, particularly anything involving legible text in the image. Strong at multi-turn editing (“now change the rug, now move the lamp”), strong at preserving the unedited regions, free in the Gemini app at a daily quota, paid for higher use. The closest thing in 2026 to “ask in plain English and get the edit you asked for”. A separate Nano Banana 2 model launched in February 2026 trades some of the Pro reasoning depth for Flash-tier speed.

OpenAI GPT Image 2 is the parallel from OpenAI, available through the API and inside ChatGPT. Best-in-class pixel stability outside the edited region (an edit changes only what you asked, leaving the rest untouched), strong text rendering, accepts up to 16 reference images for context, supports multi-turn editing. The leading choice when you are already in the OpenAI ecosystem or need API access for an integration.

Adobe Firefly is the leader for commercially safe output. Firefly is trained exclusively on licensed Adobe Stock content and public-domain material, and Adobe indemnifies paying business subscribers against copyright claims arising from generated content. Inside Photoshop, Firefly powers Generative Fill, the workhorse inpainting tool that most working creatives now reach for first. If the question is “can I sell this in a campaign without a copyright argument”, Firefly is the lowest-risk answer.

Midjourney has historically been a text-to-image tool with limited image-to-image control. The 2026 versions add reference-image and style-reference support, useful for carrying a look across a series, less suited for tight editing of an existing photo.

Specialist editing models

Black Forest Labs’ Flux Kontext ([Pro], [Max] and open-source [Dev]) is the standout for consistency in image-to-image work. Built specifically for contextual editing, it holds character, finish and identity across multiple edits more reliably than diffusion-only pipelines. Used heavily by people producing series of consistent images of the same person, the same product or the same building across different shots and angles.

Runway ML sits at the image-to-video boundary. Its image-to-image capabilities (style transfer, generative editing, structure-guided generation) are competent, but its centre of gravity is animation. Reach for it when the still is one frame in a moving deliverable.

Open-source

Stable Diffusion and the open Flux family are the centre of the open-source ecosystem. Stable Diffusion 3.5 plus the broader Flux releases provide the foundation that the rest of the community (LoRAs, ControlNet variants, fine-tunes, ComfyUI workflows) is built on. The upside is total control: run locally on your own hardware, no per-image fees, no rate limits, no terms-of-service surprises, granular access to every parameter. The downside is workflow complexity: a respectable open-source img2img setup takes hours to configure and a capable GPU to run at speed.

ControlNet is the layer most serious open-source workflows add on top. It conditions the diffusion process on explicit structural inputs (line drawings, depth maps, pose skeletons, segmentation masks), so the model preserves geometry, perspective or pose with far more reliability than a strength-dial alone. This is what makes the open stack viable for architectural sketch-to-render and any other task where structure cannot drift.

Property and design verticals

A separate cluster of tools wraps these models in property-specific workflows. Virtual-staging platforms (Virtual Staging AI, ApplyDesign, Styldod, Box Brownie, reIMG and many others) handle the empty-room-to-furnished job at volume with templates, brand-fit options and per-image pricing from around 50 cents. Architecture-specific tools (mnml.ai, Gendo, Archsynth, Vizcom, PromeAI) do the same for sketch-to-render, with structural lock built in. Houzz Pro added AI-powered finish application to its 3D Floor Plan tool in May 2026, letting professionals clip any finish from a photo and apply it across walls, floors, splashbacks and countertops in a plan view. The vertical tools win on workflow integration and consistency. The generalist tools win on flexibility and price.

The honest summary: if you are testing the category for the first time, open the Gemini app or ChatGPT and try a real edit on a real photo. The frontier models are good enough now that the friction of standing up an open-source stack is only worth it if you need volume, control or full data privacy.

Before After

Writing an image-to-image prompt that works

The same prompt-writing instincts you use for text-to-image transfer poorly to image editing. The frontier models in 2026 reward instruction-style prompting, not the keyword-stuffed comma-separated lists that worked on early Stable Diffusion. A few rules that consistently produce better edits.

Tell the model what to change and what to leave alone. “Change the splashback tile to white subway, keep the cabinetry, benchtop and floor unchanged” produces a cleaner edit than “white subway tile splashback”. The frontier models will respect the instruction.

Anchor the topic word. On an interior edit, include the room type (“this kitchen”, “this bedroom”) and the desired state (“renovated”, “staged”, “modernised”) in the prompt itself. Leaving the model to infer the subject from the image alone increases drift.

Use a reference image where possible. Most of the leading 2026 tools (GPT Image 2 up to 16 references, Nano Banana Pro, Firefly, Flux Kontext) accept a second image alongside the input as a style or material reference. “Apply the cabinetry style from the attached reference to this kitchen” is more reliable than describing the cabinetry in words.

Iterate, do not rebuild. When the first edit is 80% there, edit the result, do not start again from the original. Multi-turn editing keeps the unedited regions stable across the conversation. Restarting throws that away.

Pick the right tool for the test. Text inside the image (price tags, signs, brand wordmarks): Nano Banana Pro or GPT Image 2. Surgical removal or replacement of a specific area: Photoshop Generative Fill (Firefly) for the masking precision. Consistent character or product across multiple shots: Flux Kontext. Local control and no per-image fee: a Stable Diffusion or Flux open-source setup with ControlNet.

Lower the denoising strength when structure matters. For property and product work where the geometry of the input must hold, run the model at 0.3 to 0.5, not 0.7+. The frontier closed models hide this dial behind plain-English controls (“keep the original mostly intact” versus “creative reinterpretation”), but the underlying mechanic is the same.

Where image-to-image AI still falls down

Six places to know about in 2026 before you bet a deadline on this technology.

Text inside the image remains inconsistent on most models. Nano Banana Pro and GPT Image 2 finally cracked this in late 2025; everything older still produces gibberish on signage, price tags, screens and labels. If your edit involves readable text, use a frontier model or composite the text in post.

Exact geometric accuracy is not where these models live. A floor plan generated by image-to-image is decorative, not measured. A council photomontage that depends on accurate height, setback and overshadowing claims still requires a CAD-true 3D pipeline. AI tools can post-process the renders; they cannot replace the underlying geometry.

Identity consistency without help. Asking a generalist model for “the same person, now in a different outfit” without a reference image rarely returns the same person. The frontier multimodal models do better with explicit reference inputs; specialist tools like Flux Kontext do better again. The era of consistent character work without setup is not quite here yet.

Fine-grained material truth. A model might give you a beautiful render that looks like marble, but it is not your supplier’s marble. For client-specification work where the actual material has to appear, AI restyling is fine for concept and direction. Locked-down FF&E visuals need either a reference photo of the real material as a strong conditioning input or a traditional CGI pipeline.

Long-tail accuracy. The models are trained on the visual world as it is photographed and indexed, which biases their output. They produce convincing generic Hamptons, generic Japandi, generic Australian coastal. They are less reliable on niche or hyper-local styles, on specific Australian native plants and on regional vernacular architecture without explicit reference imagery.

Ethics and disclosure. The same capability that virtually stages an empty rental can hide damp, water damage and cracking. The same capability that day-to-dusks a beautiful exterior can move the sun in a way that conceals a north-facing problem. The Australian regulatory line has firmed in the past eighteen months and is covered next.

Before After

Image-to-image AI in Australian property and design

The largest commercial use of image-to-image AI in Australia in 2026 sits in property and home design. The specifics matter because the legal and consumer expectations here are different from the US-dominated content that ranks for most generic search terms.

Virtual staging of vacant properties is now standard practice across REA and Domain listings. Specialised virtual-staging services and done-for-you platforms produce the work at $5 to $50 per image depending on tier. AI-only platforms produce it for cents per image, with the trade-off being lower consistency on complex spaces. The standard disclosure is to tag the image as virtually staged in the listing copy; both REA and Domain have allowed this since well before AI virtual staging arrived. Our home staging guide and virtual staging guide cover the staging side in depth.

Renovation visualisation is the second-largest category, used by builders, kitchen and bathroom specialists, painters, tilers and trades to show a client what the finished job will look like before they commit. The job is image-to-image at its most literal: input photo, output the same place finished. The closest existing demand pattern, kitchen renovation visualisation and bathroom renovation visualisation, are both pure image-to-image work.

Architectural concept rendering for early-stage design conversations is now reachable by interior designers and architects without a CGI studio. Sketch-to-render workflows turn elevations, line drawings or massing models into client-presentation images in minutes. For DA submissions and council photomontages, the accuracy bar requires conventional 3D plus AI post, not AI alone.

Real-estate photo enhancement (day-to-dusk, sky replacement, lawn greening, clutter removal) was already in widespread use before AI image-to-image arrived. The change is that the work that used to require a Photoshop specialist now takes ten seconds.

The 2025 NSW disclosure rule, the Residential Tenancies (Protection of Personal Information) Amendment Bill, draws the legal line. It requires landlords and agents to disclose when rental images have been altered to “conceal faults” or “mislead rental applicants”, with examples including the removal of background infrastructure (powerlines, towers) to obscure views, the use of AI to hide damp or damage, and the placement of furniture that misrepresents the usable size of a room. Penalties are $5,500 for individuals and $22,000 for corporations. Basic editing, such as cropping and brightening, is exempt. The practical reading: virtual staging of empty rooms and disclosed cosmetic enhancement are fine, concealing faults is not. Expect other states to follow.

The broader, sometimes unstated rule across Australian property advertising is the Australian Consumer Law. Misleading or deceptive conduct in trade and commerce is illegal regardless of whether AI is involved. A virtually staged photo of an empty room is not misleading because no one expects the furniture to come with the property. A virtually de-damped photo of a wall that actually has rising damp is misleading whether you did it with AI or a human in Photoshop.

Done-for-you versus do-it-yourself

The image-to-image market splits cleanly into two camps in 2026, and the right choice depends on volume and time.

DIY means opening a tool yourself: Gemini, ChatGPT, Firefly, Stable Diffusion, your choice. Marginal cost per image is close to zero, control is total, time investment is real. Prompt iteration alone for a non-trivial property edit can run 30 minutes to an hour for someone learning the tool. Watermark stripping, format conversion, batch processing, consistency across a campaign and quality control all become your problem. Worth it when you have time, taste and a small enough job to absorb the friction.

Done-for-you means submitting a photo and a brief and getting the result back. The category includes specialist virtual-staging services, sketch-to-render studios, AU-based marketing-focused services like reIMG, and traditional CGI studios that have absorbed AI tooling into their pipeline. The marginal cost per image is higher, the time investment from you is near zero, the consistency and quality control are someone else’s job. Worth it for the property that matters, the quote you really want to win, the campaign with deadline pressure, the volume you cannot service in-house.

The break-even is usually about quantity and stakes. One important image per month: DIY is fine, even fun. Twenty images a month against listing deadlines: done-for-you pays for itself in the time you don’t spend learning the latest prompting quirks. For the deeper economics, how virtual staging is actually priced is a useful read.

What changes next

The trajectory through 2026 is converging. The big multimodal models (Gemini, GPT) are absorbing what used to be separate jobs (text-to-image, image-to-image, video) into a single interface. The specialist models (Firefly, Flux Kontext, Runway, the vertical staging and architectural rendering tools) are differentiating on commercial safety, identity consistency and workflow integration rather than on raw quality. Open-source remains the place for control and self-hosting, with Stable Diffusion 3.5 and the open Flux family as the foundation.

Two practical predictions worth holding. First, native image-to-image inside the design tools people already use (Photoshop, Illustrator, Figma, Canva) will become the default surface for the work, not the model-vendor apps. Most professionals will reach for the model from inside the tool they edit in, not the other way around. Second, the legal regime will continue to firm. NSW’s 2025 disclosure rule is the first formal Australian framework specifically targeting AI-altered listings; expect Victoria, Queensland and federal Australian Consumer Law guidance to follow, and assume any commercial use of AI image editing in a public-facing context should be disclosable, traceable and defensible. The technology stays useful. The expectations around using it openly will only get firmer.

For the practical jobs this site exists to do, the upshot is small and concrete. Image-to-image AI changed the cost and the speed of producing the imagery property professionals, trades and designers needed anyway. It did not change what the imagery is for, which is helping a client see what something will look like clearly enough to say yes.

Frequently asked questions

What is image-to-image AI in simple terms?

Image-to-image AI takes an existing picture and an instruction, and gives you back a new picture that keeps the original’s structure and composition but applies the change you asked for. The classic example is feeding in a photo of an empty room and a prompt like “add modern coastal furniture” to get back the same room, same windows, same floor, now furnished. It is the opposite of text-to-image, which starts from a blank canvas and a description and invents everything from scratch.

What’s the difference between image-to-image AI and text-to-image AI?

Text-to-image starts with words and a blank canvas, so it invents the scene, the camera angle, the lighting, the geometry, all of it. Image-to-image starts with an actual photo or a 3D render and changes only what you ask it to change, holding the rest in place. For anything that has to match a real room, a real product, a real building or a real listing, image-to-image is the relevant approach. Text-to-image is the relevant one when you’re inventing something that doesn’t exist yet.

Which is the best image-to-image AI tool in 2026?

There is no single best. For everyday editing inside an existing creative workflow, Google’s Nano Banana Pro (the Gemini 3 image model) and OpenAI’s GPT Image 2 currently lead on edit fidelity and instruction-following. Adobe Firefly leads on commercial-safe output because of its licensed training data and indemnification. Black Forest Labs’ Flux Kontext leads on character and scene consistency across edits. Stable Diffusion and the open Flux models remain the leaders for full local control. The right pick depends on how much control, how much legal certainty and how much commercial use you need.

Is image-to-image AI free?

Most of the leading models have a free tier that is enough to try them and produce occasional images. Google’s Gemini app gives free users a daily quota of Nano Banana edits. Adobe Firefly’s web app offers a small monthly generative-credit allowance. Stable Diffusion and the open Flux models are completely free if you run them on your own computer, though you need a capable GPU. Heavy or commercial use almost always pushes you onto a paid plan, where pricing sits in the A$25 to A$80 per month range depending on the tool and quota.

Can I sell or publish images made with image-to-image AI?

Sometimes. Adobe Firefly is the safest because Adobe trained it on licensed data and indemnifies paying business customers against copyright claims. Most other major tools allow commercial use under their terms but do not indemnify you, which means you carry any infringement risk yourself. Some open-source models have licences that restrict commercial use entirely. Read the licence for whichever tool you use, and for any real-estate, brand or client-facing work, prefer tools with a clear commercial-use stance.

Is image-to-image AI on Australian real estate listings legal?

Yes, with disclosure becoming the line. Editing photos has always been allowed in Australian real-estate marketing, but using AI to remove faults, hide damage, or fabricate features that the property does not have is misleading conduct. NSW introduced a 2025 bill requiring landlords and agents to disclose AI-altered rental images, with penalties of $5,500 for individuals and $22,000 for corporations on non-disclosure. The practical rule: virtual staging of empty rooms and obvious cosmetic enhancement are fine if disclosed. Concealing faults is not.

What does denoising strength mean?

Denoising strength is the dial on most image-to-image models that controls how much the AI is allowed to change the original. It runs from 0 to 1. Low values, around 0.2 to 0.4, keep the original almost intact and only nudge style or detail. Mid values, around 0.5 to 0.7, allow real changes while keeping the composition recognisable. High values, above 0.8, approach text-to-image and effectively rebuild the picture from scratch. For property and product work where structure has to hold, stay in the lower half of the range.

Where does image-to-image AI still fall down in 2026?

Three places. First, text inside images: signs, labels, price tags and brand wordmarks are still inconsistent except on the very latest models like Nano Banana Pro and GPT Image 2. Second, exact geometric or proportional accuracy: floor plans, council-grade photomontages and engineering drawings still need conventional 3D, not AI. Third, identity and brand consistency without help: keeping the same person, the same product or the same building looking like itself across multiple images requires reference tools, LoRAs or models like Flux Kontext built specifically for the job.