The first image from a generator is usually the easy part. The trouble starts with image four, when the mug that was sage green in the hero shot comes back teal, the warm window light turns into a cold studio glow, and someone notices the handle now has a second, smaller handle growing out of it. Side by side, the set looks like it came from four different businesses.
AI image consistency is the work of keeping a set of generated images visually related: same palette, same light, same materials, same way of framing a subject. This article is the execution side of that problem. It assumes your brand already has approved visual rules (if it doesn’t, start with Direktorium’s piece on locking a visual system into your brand prompts) and shows how to turn those rules into a specification, generate a set, catch the image that drifted, and fix it without starting over.
Know which kind of consistency you need
People say “consistent” and mean four different things, and generators handle them very differently.
| What you want | How far AI gets you | What to do instead or as well |
|---|---|---|
| Stylistic consistency: palette, light, mood, framing | Reliable with good references and a written spec | Review every image against the spec |
| Exact product fidelity: your real product, every detail right | Unreliable; shapes, logos and proportions drift | Use real product photos and composite them in |
| Reliable text: prices, slogans, logo lettering | Improved, still error-prone | Add all text in a design tool afterwards |
| Reproducible identity: the same person or character every time | Possible with reference features, never guaranteed | Use real photos of consenting people for anything that matters |
The vendors say the same thing in their own documentation. OpenAI lists consistency as a known limitation of its image models, noting they “may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations,” and that text placement can still go wrong (OpenAI image generation guide). Its prompting guide goes further: if a region must stay pixel-identical, composite the approved edit into the original image instead of relying on prompting alone (OpenAI image prompting guide). For product shots, generate the scene and place the real product photograph into it.
The same logic applies to your logo. A generated image with a logo-ish shape on a bag is a concept. Your editable brand masters (vector logo, exact colour values, licensed fonts) stay the source of truth, and the logo goes on in the layout stage. The brand launch checklist covers how those masters get produced and handed over, and the agritech logo article deals with what a mark should say in the first place. Neither job belongs inside an image prompt.
Build the art-direction spec before generating anything
Most drift comes from prompts written fresh for each image: “cosy autumn kitchen, warm light” for image one, “autumn table scene, soft natural light” for image three, and the model reasonably treats them as two different requests. A specification fixes what must not change and leaves one named slot for the scene.
Fill this template once per campaign and paste from it every time.
| Field | What to write | Example entry |
|---|---|---|
| Campaign and use | Name, channels, number of images | Autumn range launch: website hero, Instagram feed, Story, email banner |
| Approved references | File names and what each one governs | REF-A governs light and palette; REF-B governs composition; REF-C is the real product photo |
| Composition rules | Camera height, distance, subject position, negative space | Eye level with tabletop, product in left third, top quarter kept clear for text |
| Palette | Named colours with values, and where they may appear | Sage glaze #8A9A7B on product only; oat linen #E8DFD0 surfaces; no saturated blues |
| Lighting | Direction, quality, time of day, colour temperature | Single window light from left, late afternoon, warm, soft shadows to the right |
| Materials and surfaces | Textures that recur across the set | Unvarnished oak, washed linen, matte stoneware |
| Subject details (locked) | Everything about the product or character that must never change | One handle, speckled glaze, unglazed foot ring, no printed logo |
| Exclusions | What must never appear | Text, logos, hands with visible faces, steam swirls, plastic, cool or blue light |
| Aspect ratios and safe zones | Each format and the area kept clear for layout | 4:5 feed, 9:16 Story (top and bottom 15% clear), 16:9 hero, 3:1 email banner |
| Tool controls used | Reference images, reference strength, seeds, model version | Style reference REF-A at default strength; model version noted in the run log |
| Reviewer and sign-off | Who approves, against which fields | Owner approves palette and product; designer checks composition and safe zones |
“Approved references” turns taste into files: agreed images beat a paragraph of adjectives. “Subject details (locked)” is what your reviewer checks every image against, so it has to be specific enough to fail. “Looks nice” cannot fail. “One handle” can.
The spec then becomes a fixed prompt block with one variable line. Anything above the scene line is pasted unchanged into every generation.
Style: matte stoneware product photography, editorial and calm. Light: single soft window light from the left, late afternoon, warm, gentle shadows falling right. Palette: sage green glaze on the product only, oat linen and pale oak surfaces, no saturated blues. Materials: unvarnished oak table, washed linen, matte ceramic. Composition: camera at tabletop eye level, product in the left third, top quarter clear. Exclusions: no text, no logos, no steam, no plastic, no cool or blue light. Scene: [one sentence describing this image only]
Use the reference controls your tool actually has
The controls differ by tool and change names quickly, so check the current documentation for your version. The notes below were checked on 1 November 2026.
| Tool | Documented controls for consistency | Worth knowing |
|---|---|---|
| Midjourney | Style Reference (–sref) with style weight –sw from 0 to 1000; –seed | V8.2 became the default on 24 July 2026 and its Edit Model replaces Omni Reference, taking up to four reference images (Version notes) |
| Adobe Firefly | Style reference and composition reference, each with a strength slider | In the API, both take a strength from 1 to 100, default 50 (style, structure) |
| Gemini image models (Nano Banana) | Up to 14 reference images in the API, split by model into object, character and style references | Gemini 3 Pro Image takes up to five character images and up to three style references; all outputs carry a SynthID watermark (Gemini API docs) |
| OpenAI GPT Image / ChatGPT | One or more input images as references; multi-turn editing | The ChatGPT guide suggests numbering images and stating what each one controls (“Image 2 is the style reference”) |
A word on seeds, because they get oversold. A seed fixes the starting noise, which helps when you want to compare two prompt wordings on equal terms. Midjourney’s own documentation says seeds can’t capture or bookmark a style, character or appearance across different prompts, have the least impact on the final image, and may not behave the same if you come back later. For a campaign, that makes them a testing tool. Your consistency comes from references and the written spec, and a seed number in the run log is a nice extra.
Google’s prompt guide adds a habit that works in every tool: give each recurring object a name and use it every time (Google DeepMind prompt guide). “The Saltmarsh mug” in every prompt gives the model one thing to track. “A ceramic mug” in one prompt and “a stoneware cup” in the next invites two different objects.
A coherent set and the one that drifted
To make this concrete, here is an illustrative review sheet for a fictional ceramics studio, Saltmarsh Pottery, launching an autumn range. The business, images and outcomes are invented for this example; no real generation run is reported. Each row is one image, checked against the spec fields above.
| Image | Ratio | Composition | Palette | Light | Locked details | Verdict |
|---|---|---|---|---|---|---|
| 1. Hero, mug on oak table | 16:9 | Pass | Pass | Pass | Pass | Approved |
| 2. Feed, two mugs and bread | 4:5 | Pass | Pass | Pass | Pass | Approved |
| 3. Feed, hands holding bowl | 4:5 | Pass | Pass | Pass | Bowl glaze slightly darker | Approved with product photo composited |
| 4. Feed, shelf of the range | 4:5 | Pass | Pass | Pass | Pass | Approved |
| 5. Story, mug by window | 9:16 | Product centred, no clear top zone | Glaze shifted to teal | Cool, overcast | Two handles; invented logo on mug | Rejected |
| 6. Email banner, table edge | 3:1 | Pass | Pass | Pass | Pass | Approved |
Image 5 is the failed variation, and the reason is visible in the run log: whoever generated it wrote a fresh prompt (“cosy mug by a rainy window, vertical”), dropped the style reference, and let the scene line overrule the light. Rain implies overcast light, so the model obliged. The extra handle and the made-up logo are the kind of product drift the table at the top warned about. The model did what the prompt asked.
The fix had three parts. The style reference went back in. The scene line was rewritten so it could not compete with the light rule (“the Saltmarsh mug on a windowsill, late afternoon sun, dry day”). And the mug itself was replaced with the studio’s real product photo in the edit, because a fourth attempt at getting one handle right is a worse use of time than a ten-minute composite.
Image 3 was the softer case: the glaze was a shade off, everything else held, so the real bowl was composited in rather than risking a good composition on another attempt.
Changing aspect ratio without losing the look
Cropping a 16:9 hero into a 9:16 Story rarely works: you lose the product or the negative space. Generating each format as its own image works better, as long as everything except the scene and the ratio stays fixed. Most tools let you set the ratio directly (Midjourney’s –ar parameter, the aspect_ratio setting in the Gemini API), so add a ratio line to the prompt block and a safe-zone rule for each format.
Composition rules are usually written for landscape, so the spec should say what changes for vertical formats: for 9:16, product in the lower middle, top and bottom 15% clear for platform interface and text.
A short review routine
Review the set together at thumbnail size, which is how most people will see it. Drift that hides at full size (a slightly cooler white balance, a heavier shadow) jumps out in a grid.
- Check each image against the locked subject details first. A wrong product is an automatic reject.
- Check palette and light against the references, side by side.
- Check composition and safe zones per format.
- Decide per image: approve, composite a real element, or regenerate with the spec unchanged and the scene line clarified.
- Log the prompt block, references, tool, model version and any seed for every approved image.
Six weeks later someone will need a seventh image, and the run log is the difference between a match and a lookalike. If you plan campaigns ahead, an AI content calendar tells you how many images each campaign needs before you start generating.
Where the spec lives
An art-direction spec is campaign-specific, but most of it comes from rules your brand already has. Palette, materials and exclusions should be lifted straight from your AI brand guidelines, so the image prompts and the copy prompts follow the same source. The pet-care branding guide shows how a visual identity direction gets set for a small business; this workflow is what happens after that direction is approved.
Expect a coherent set and a few composites, never perfection. Direktorium has more practical workflows for small teams using AI in production. If you would rather hand campaign visuals to people who do this every week, BBDirector can help.
