Consistency comes from process, not the model: a written style recipe, reference images, plain backgrounds and a human review, then a real web pipeline with AVIF or WebP, responsive sizes and an eagerly loaded hero. Check usage rights and India's proposed AI labelling rules before you publish.
You can make brand-consistent assets with AI image generators, but the consistency comes from process, not from the model. Write a fixed style recipe, give the generator reference images as well as words, generate on plain backgrounds, review every image by eye, and send the keepers through a proper web image pipeline. Skip those steps and you get a folder of attractive pictures that do not look like one brand.
The tools moved quickly in 2025. Google released Gemini 2.5 Flash Image in August, built to keep a subject consistent across edits, at about US$0.039 per image through its API. Adobe put Firefly Image Model 5 into public beta on 28 October, generating at a native 4 megapixels. Generation is now the cheap step. Quality is decided by everything around it.
- per image for Gemini 2.5 Flash Image through Google's API
- US$0.039
- per image for Gemini 2.5 Flash Image through Google's API
- of mobile pages have an image as their Largest Contentful Paint element (Web Almanac 2024)
- 73%
- of mobile pages have an image as their Largest Contentful Paint element (Web Almanac 2024)
- of mobile pages with an image LCP lazy-load it, which delays the paint
- 16%
- of mobile pages with an image LCP lazy-load it, which delays the paint
- of pages served their LCP image as AVIF in 2024, while JPG led with 61% on desktop
- 0.3%
- of pages served their LCP image as AVIF in 2024, while JPG led with 61% on desktop
01
Write the style recipe before the first prompt
Treat the prompt as a specification that every asset in the set shares word for word. Change it and you have started a new set. Keep it in a shared document with three approved reference images, and make everyone who generates for the brand start there. A useful recipe has five parts:
- Palette: brand colours named in words, plus a swatch image where the tool accepts one. Treat hex codes in a prompt as a hint, not a guarantee, and correct colours afterwards.
- Rendering language: one medium and finish, such as 'soft 3D clay, matte surfaces, rounded edges', never mixed with 'flat vector' in the same set.
- Lighting and camera: one setup, for example 'soft studio light from the top left, three-quarter view'.
- Composition: subject centred, generous margins, plain background, no text in the image.
- Exclusions: what must never appear, from extra fingers to a competitor's colours.
02
Give the model references, not just words
Words drift; reference images give the model something concrete to match, and current models are built to use them. Google says Gemini 2.5 Flash Image keeps a subject's appearance consistent across prompts and edits, and can merge several input images, for instance placing your product in a new scene. At its MAX conference in October, Adobe announced Firefly Custom Models in private beta: creators train a private model on images they have the rights to use, so a whole series shares one style.
Even so, OpenAI's image-generation guide cautions that its model may occasionally struggle to keep recurring characters or brand elements consistent across generations. Generate several candidates for each asset and keep only the one that matches the set.
03
Generate clean, cut out, and upscale only what needs it
Generate subjects on plain or transparent backgrounds so each asset works on a card, a hero banner or a dark section. OpenAI's gpt-image-1 model, for example, can return a transparent background when the output is PNG or WebP. Where a tool cannot, remove the background in an editor and check the edges at full size.
Upscale only the pieces that will display large, such as heroes and banners; Firefly now lists Topaz Bloom among its partner models for upscaling. Keep a full-resolution master of every approved asset, filed with the prompt, references and model that produced it, so you can extend the set next year.
04
Ship it through a real image pipeline
A beautiful 3 MB hero image is still a slow page. Images are the Largest Contentful Paint element on 73.3% of mobile pages, according to the 2024 Web Almanac, and LCP is a Core Web Vital: Google's web.dev guidance asks for 2.5 seconds or less for at least 75% of page visits. It recommends AVIF or WebP, responsive sizes, compression, and never lazy-loading the LCP image.
| Asset | Keep as master | Serve on the site as |
|---|---|---|
| Hero illustration or banner | Full-size PNG, upscaled | AVIF or WebP at several widths via srcset and sizes, loaded eagerly with fetchpriority="high" |
| Cut-out product or character | Transparent PNG | WebP with transparency, width and height set |
| Icons and simple marks | Vector redrawn by a designer | SVG |
| Images below the fold | Full-size PNG | AVIF or WebP, lazy-loaded |
Two more rules prevent common mistakes. Never lazy-load the image likely to be the LCP element, as 16% of mobile pages with an image LCP did in 2024. And give every image a width and height: web.dev lists images with unknown dimensions among the causes of layout shift. Frameworks help. Next.js serves WebP by default and can add AVIF, which its documentation says compresses about 20% smaller than WebP but takes about 50% longer to encode.
Name files descriptively and write alt text that says what the image shows; Google's image guidelines call alt text the most important attribute for image metadata. Purely decorative illustrations should carry an empty alt attribute, as W3C's accessibility tutorials advise, so screen readers skip them.
05
Rights, labels and disclosure
Ownership of purely generated images is unsettled. In January 2025 the US Copyright Office concluded that generative AI outputs "can be protected by copyright only where a human author has determined sufficient expressive elements", and that prompts alone are not enough, though human arrangement or modification of the output can be protected. Keep your logo and core brand marks human-made, and keep records of the human work in each asset. Read each tool's commercial terms too; Adobe, for one, describes its Firefly models as commercially safe.
India has proposed mandatory labels. On 22 October 2025, MeitY published draft amendments to the IT Rules that would require AI-generated visuals to carry a label covering at least 10% of the screen area, and would make large social platforms verify users' declarations about synthetic content. Comments are open until 13 November. Meta already shows an 'AI info' label on Facebook and Instagram when it detects industry-standard AI signals or when people disclose AI content.
06
Review every asset before it ships
- Hands, fingers, eyes and teeth, checked at full size.
- Text and lettering removed or regenerated, then real type added in your design tool.
- Logos, labels and products shown exactly, or not at all.
- Cultural details right for your audience, from a sadya served properly to a mundu worn the right way.
- Colours checked against the palette with a colour picker, not judged by eye on one screen.
- Weight, dimensions, filename and alt text confirmed before upload.
What to do this week
- Write a one-page style recipe and choose three reference images.
- Generate one small set, say six spot illustrations, and review it against the checklist.
- Set up the pipeline once: masters, AVIF or WebP exports, srcset sizes and an eagerly loaded hero.
- Record the prompt, references and model with every approved asset.
- Decide now how you will label AI imagery, so India's final rules do not catch you out.
Sources
Facts current as of 7 November 2025.