AI image editing needs a production workflow

Website imagery often becomes inconsistent one small request at a time.

A product needs a cleaner background. A campaign needs a different crop. An old service image no longer matches the page. A designer fixes the urgent asset, but the wider library keeps accumulating mixed dimensions, styles and levels of quality.

More capable AI image editing can reduce some of that repeated production work. It does not make every generated image safe to publish.

The useful opportunity is a controlled asset workflow: begin with approved source material, define what must remain accurate, test a representative batch, keep human approval and optimise the final files for the website.

What changed on 8 September 2026

OpenAI released ChatGPT Images 2.5 on 8 September. It began rolling out across all ChatGPT, ChatGPT Work and Codex tiers on desktop, mobile and web that day.

OpenAI also made two models available through its API. GPT Image 2.5 Flare is intended for faster everyday generation and editing. GPT Image 2.5 Sunburst is intended for work where editing precision matters more than speed.

OpenAI reports better preservation of reference subjects, more precise local changes, stronger consistency through several edits and up to 50 per cent lower latency than Images 2.0. These are vendor claims, not a guarantee that a particular product, brand or website workflow will meet its acceptance standard.

The practical capability is still meaningful.

The Image API documentation supports image edits, transparent backgrounds, custom dimensions, PNG, JPEG and WebP output, and adjustable quality. It also documents important limits: text can still be rendered incorrectly, recurring brand elements can drift and structured compositions may not place elements precisely.

That combination makes the development useful enough to test, but not reliable enough to remove review.

Why this matters to a smaller business

Many small and medium businesses do not need a new image generation platform.

They need a more consistent way to maintain the images already supporting product pages, service pages, campaigns, articles and ecommerce catalogues.

A controlled editing workflow could help with:

• removing or replacing backgrounds around an existing product;
• creating approved crops and dimensions for different page placements;
• adapting a campaign asset while preserving its main subject;
• preparing transparent cutouts for catalogues or comparison pages;
• extending a real image to suit a different aspect ratio;
• producing draft variations before a designer or business owner selects the final direction.

This is a credible AI Website Upgrade when the current website still works and the asset process is the constraint. It can improve the existing content layer without changing the CMS, commerce platform or page architecture.

The business case should still be measured against the current process. Faster generation has little value if staff spend more time correcting packaging, checking text or rejecting images that do not look like the real product.

Start with the truth the image must preserve

Before choosing a prompt or model, define which parts of the source image are factual.

For a product image, that may include its shape, colour, material, label, controls, accessories, scale and packaging. For a service business, it may include the real person, location, equipment or work being represented. For a brand asset, it may include the logo, typography, approved colours and composition.

These are not creative preferences. They are acceptance criteria.

The Australian Competition and Consumer Commission says images and descriptions used to promote goods or services must be accurate and truthful. A synthetic product image can therefore create a commercial problem if it shows a feature, finish, result or context that the customer will not receive.

Human review should compare the output with the source and the real offer, not only ask whether it looks polished.

Run a ten asset pilot before changing the workflow

Choose ten assets that represent the difficult parts of the library.

Include simple and complex edges, reflective or transparent materials, small text, branded packaging, people where appropriate, and the aspect ratios the website actually needs. Avoid selecting only the cleanest source photographs.

Before the pilot, record:

• the time currently spent preparing each asset;
• the number of review and correction rounds;
• the proportion of assets approved for publication;
• the final file weight and dimensions;
• any engagement or conversion measure the page already uses;
• the common reasons an asset is rejected.

Then give the model a bounded role. Ask it to perform one defined edit from an approved original. Do not ask it to reinvent product details that the source does not show.

Use low quality output for early direction where appropriate, then compare higher quality output only after the edit has been accepted. Test Flare and Sunburst on the same difficult examples rather than assuming the premium workflow is necessary for every asset.

The pilot should answer whether the complete process is better, not whether the model can produce one impressive image.

Use an acceptance checklist for every final asset

An approval step becomes useful when the reviewer knows what to check.

For each proposed image, verify:

• factual product or service details match the approved source;
• logos, labels and written copy are correct;
• colours and materials are not misleading;
• people, locations and sensitive details have been handled appropriately;
• the image fits the intended page and crop without hiding important content;
• the final file uses an appropriate format, compression and responsive dimensions;
• alternative text describes the image's purpose on the page;
• the business has the rights and consent needed for source and output;
• the approved original and final asset can be found again later.

Consequential assets should remain human approved. That includes product images, before and after comparisons, regulated claims, real people, customer work and any visual used as evidence of an outcome.

Keep privacy and rights decisions outside the prompt

An upload can contain personal information, client material, location details or protected creative work.

The Office of the Australian Information Commissioner recommends due diligence before using commercially available AI products, including reviewing who can access input and generated information, how human oversight works and whether personal information is necessary for the task. The OAIC also notes that generated or inferred images about an identifiable person can be personal information.

Use approved business accounts and settings rather than a staff member's unmanaged personal account. Minimise what is uploaded. Obtain appropriate privacy or legal advice where the business's obligations or consent basis are unclear.

Rights need similar care.

IP Australia advises businesses to review what protected material is being supplied to an AI tool, the tool's ownership and commercial use terms, and whether the output could conflict with existing intellectual property.

OpenAI's current Terms of Use assign its rights in output to the user to the extent permitted by law, but also say output may not be unique and require the user to hold the necessary rights for input. Business and API accounts may operate under a different agreement, so the contract applying to the account should be checked. OpenAI's service terms for image capabilities also require express consent and necessary rights when reproducing a person's likeness.

Those terms are inputs to the decision, not a substitute for checking the actual source material and intended use.

Provenance helps, but it does not prove accuracy

OpenAI says supported images generated through ChatGPT, Codex and its API include C2PA metadata and SynthID watermarking.

Its provenance guidance also explains the limitation. Metadata can be removed by editing, conversion or publishing platforms. A detected signal can indicate that OpenAI generated or exported the file, but it does not confirm that the image is accurate, legally owned, unedited or presented in the right context.

Preserve provenance signals where the publishing workflow allows it. Keep a separate production record as well.

For each approved asset, retain the original, the final file, the intended use, the model or product used, the main edit instruction, the reviewer and the approval date. If the API is part of a repeatable process, consider using the dated model snapshot during evaluation so a model change does not silently alter the result halfway through the batch.

This record also reduces vendor dependence. The business can move the workflow later without losing the source assets, acceptance rules and history behind published images.

Website optimisation still happens after generation

A generated image is not automatically a production website asset.

It still needs the correct crop, dimensions, compression, responsive behaviour and alternative text. It needs to be placed where it helps a customer understand the offer. It needs to avoid unnecessary page weight and layout shift.

The API's published prices for both GPT Image 2.5 models are US$5 per million text input tokens, US$8 per million image input tokens and US$30 per million image output tokens. Actual workflow cost depends on the number and size of source images, output dimensions, quality settings and correction rounds.

Measure the human cost as well as the API cost.

A workflow that generates more options but creates a larger approval queue is not yet an improvement. A workflow that produces a smaller number of accurate, reusable assets may be valuable even when its model bill is not the lowest.

Ongoing Website Growth & Care should review whether the new assets improve the live experience, remain factually current and continue to earn their page weight.

The next decision

Better image models make controlled editing more practical. They do not turn website imagery into an unattended automation task.

Start with one real asset problem and ten representative files. Establish the current baseline. Define what the model may change and what it must preserve. Keep human approval at the point of publication. Measure correction effort, acceptance rate, performance and customer response before expanding the workflow.

If the current website has useful content but its imagery is inconsistent or expensive to maintain, AI Website Upgrade can test the smallest useful improvement without forcing a rebuild.

When the decision spans brand, website performance, privacy, production tooling and future maintenance, a Fractional Technical Partner can help define the workflow before the business commits to a larger platform or asset programme.