For repeatable on-brand images, Adobe Firefly is easier to control: a style reference with a Strength slider on Adobe's own Firefly Image models, model choice, and outputs Adobe calls safe for commercial use. ChatGPT Images is easier to steer by conversation, saving a reference photo and editing one area at a time. Explore in ChatGPT; produce in Firefly.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowThe detail that changes the answer is which model you're using inside Firefly. Adobe documents the reference-plus-Strength control for its Firefly Image models; switch to a partner model such as GPT Image and you're working with reference images and prompts instead, paying credits, and outside Adobe's commercial-safety statement. Meanwhile both tools have moved towards each other: Firefly now has a conversational AI Assistant in beta, and ChatGPT can save a reference photo to reuse by name.
What does on-brand mean for an AI image?
An on-brand image is one that passes six tests you can mark pass or fail: palette, look, recurring subject, layout room, local edits and no invented marks. "On-brand" is too vague to compare tools against, so these are the tests used throughout this tutorial:
- Palette. The dominant colours sit within your brand range. A forest-green brand shouldn't come back teal.
- Look. The lighting, finish and level of polish match your existing photography, so a new image doesn't stand out on your feed.
- Recurring subject. Things that should look the same each time (a uniform, a vehicle, a product) do.
- Layout room. There's clear space where your headline and logo will go.
- Local edits. You can change one thing (a shirt colour, a box label) without the rest of the image shifting.
- No invented marks. No fake logos, lettering or number plates appear.
The sixth test is the one neither tool reliably passes on its own, which is why every workflow below ends in a layout app with your real logo file. For a fuller treatment of palettes, style words and limits, see brand guidelines for AI images.
How does Adobe Firefly keep images on-brand?
Adobe Firefly keeps images on-brand mainly through settings rather than sentences: a style reference with a Strength slider, effects and a composition reference. On the Firefly homepage you choose Image, then Generate image, type a prompt, and under the Styles section add a reference image from your device or Adobe's gallery. Adobe's help page says adding a reference activates the Strength slider: slide right and new images follow the reference more closely, slide left and they drift further from it. Below that, effects for colour tone, lighting and camera angle refine the look, and a separate composition reference can copy the arrangement of another image.
That combination is what makes Firefly good at the "look" test. Once you've found one image that feels like your brand, it becomes the reference for every image in the campaign, and Strength becomes a dial you can write down and reuse. If your house look sits at about three-quarters on the slider, everyone on the team can set it there.
Two newer tools push further. Custom models, a beta, let you train Firefly on your own images to reproduce a style or a specific subject such as a product. Adobe's credit FAQ lists 500 credits per training and 20 per generation, and Adobe lists availability for Creative Cloud individual plans, Creative Cloud Pro and Pro Plus for teams, and some enterprise plans, so check your plan before counting on it. And Firefly's AI Assistant, also in beta, has a skill called /create-a-brand-kit, can extract colours, guidelines and logos from a brand PDF, generates with /generate-on-brand-content, and can check an image against a saved brand and suggest fixes. The assistant uses credits for most actions and chooses between Adobe and partner models unless you name one, so name a Firefly model when commercial safety matters.
For local edits, Firefly's Generative fill works on a painted selection, which keeps changes to the area you mark. The dedicated tutorial on changing a product photo background with Generative fill shows the masking steps.
How does ChatGPT keep images on-brand?
ChatGPT keeps images on-brand mainly through sentences: you describe an image, look at the result, and say what to change. OpenAI's Images in ChatGPT help page says the current version, ChatGPT Images 2.6, does a better job of keeping details consistent from one edit to the next, which is exactly what brand work needs. You can skip the selection tool and describe the change, or open the editor, highlight an area with Select and describe the change for that part. OpenAI is candid about the limit: highlights are not always precise, and edits may extend beyond the area you selected.
Three features matter most for brand work:
- Saved references. Upload a photo, ask ChatGPT to save it as a reference, name it, and call it in later chats by typing @ and the name. A photo of your real van or uniform saved as @van keeps that subject in the conversation without re-uploading.
- Projects. A project holds chats, files and instructions together, and project instructions override your general custom instructions. Put your brand rules in the project settings and every image chat inside it starts with them.
- Aspect ratio and transparency. The aspect ratio picker or a line in the prompt sets the shape, and ChatGPT can return a transparent background for cut-out elements.
OpenAI lists Images with thinking for Plus, Pro and Business. On pricing, Free has limited image creation, Go ($8 a month) more, Plus ($20) "more complex and accurate image creation" with limits that apply, and Pro (from $100) unlimited and faster creation. ChatGPT Business costs $20 a seat a month billed annually or $25 monthly, from two users.
A removals firm's brand brief, run through both
Here's the main worked example. An illustrative eight-van removals firm has a clear brand: forest-green vans with cream lettering, crews in green polo shirts, and photography that looks like a real weekday rather than an advert, with natural light, ordinary streets and no beaming stock-photo smiles. For a spring campaign it needs eight supporting images: a van parked on a residential street, two crew members carrying a wrapped sofa, labelled boxes in an empty room, packing materials laid flat, a crew member taping a box, a loaded van seen from the back, a stack of wardrobe boxes, and a tidy storage unit.
The marketing lead gave each tool two hours and scored every finished image against the six tests. The scoresheet below is illustrative, filled in as an example of what to record; your own scores will depend on your brief.
| Test (8 images) | Firefly, Firefly Image model, reference at about 75% Strength | ChatGPT Images, one chat inside a brand project |
|---|---|---|
| Palette within brand range | 7 of 8 | 6 of 8, 8 after one follow-up each |
| Consistent look across the set | 8 of 8 | 5 of 8; later images drifted warmer |
| Recurring subject (green polos) | 6 of 8 | 7 of 8 |
| Layout room for headline | 7 of 8 | 8 of 8 |
| Local edits without side effects | Precise with Generative fill | Fast, but 2 edits changed nearby areas |
| No invented marks | 5 of 8 (lettering on van sides) | 5 of 8 (number plates and box labels) |
| Average attempts per keeper | 3.1 | 2.4 including follow-up edits |
Read the pattern rather than the exact scores. Firefly held one look across eight images because the reference and Strength setting did the remembering; ChatGPT needed fewer attempts per image because each fix was a sentence, but its look wandered across a long session. Both invented marks, which every brand workflow has to catch. On cost, the trial fitted inside Firefly Standard at $9.99: with a standard Firefly Image model, image generations don't use credits on a paid plan. The ChatGPT side used the firm's existing Plus seat at $20.
The firm's decision: use ChatGPT inside a brand project to rough out scene ideas and wording with the team, then produce the final set in Firefly with the saved reference and Strength setting, and finish everything in its layout tool. The hero image on the campaign landing page stayed a real photo of a real van, because a generated van can't be relied on to show your actual livery.
The firm also priced the next step up, a Firefly custom model trained on its own crew photos, to see whether it would fix the uniform scores. At Adobe's listed rates, training costs 500 credits once and each image from the model costs 20. A month of 40 images would therefore cost 500 + (40 × 20) = 1,300 credits in the first month and 800 after, which is too much for Standard's 2,000 once anything else is added, and comfortable on a 4,000-credit plan. The firm's rule of thumb: train a custom model only when the same subject (a product, a uniform, an illustrated character) must appear in dozens of images a month, and when you have a good set of your own photos to train it on. Eight supporting images for one campaign didn't meet that bar, so the standard model plus a reference stayed the cheaper answer.
If you want to keep the scoresheet in a spreadsheet and total the scores automatically, try The Complete Excel Course with AI Integration on Coding Liquids.
One brief, two prompts, and what came back
To see the difference in working style, here's the sofa image done both ways. In Firefly, with the brand reference image attached and Strength set to about three-quarters, the prompt was short, because the reference carries the look:
Two removal crew members in plain forest-green polo shirts carrying a sofa
wrapped in clear plastic down a front path, ordinary weekday morning,
natural overcast light, residential street, empty space in the left third,
no text, no logos, no number plates
A sample result, described for illustration: the light and colour matched the rest of the set first time. One image gave a crew member a tiny invented logo on the chest, fixed with Generative fill on that patch. Another had the sofa floating slightly above their hands, so it was rejected outright.
In ChatGPT, inside the project holding the brand rules, the same image took a conversation:
Turn 1: Use @crew as the reference for uniform and style. Two crew members
carry a plastic-wrapped sofa down a front path on an overcast weekday
morning. Portrait 4:5 shape with empty space on the left. No text or logos.
Turn 2: Good. Make the polo shirts the deeper green from @crew, and remove
the writing on the box in the background. Keep everything else the same.
Turn 3: The path changed from brick to concrete. Put the brick path back.
A sample result, again illustrative: turn 1 produced a usable composition with the shirts a shade too bright. Turn 2 fixed the green and removed the writing, but the edit also re-surfaced the path, which turn 3 put right. In this illustration, three sentences got a keeper without starting again. The cost is that every turn is a chance for something else to shift, so the final check matters more. The prompt patterns that keep ChatGPT steady over a session are covered in ChatGPT product photoshoot prompts for backgrounds and consistency.
Where does each tool drift off-brand, and how do you fix it?
Firefly most often drifts when a team switches models mid-campaign, sets Strength too high or lets the assistant choose the model. ChatGPT drifts over long sessions and when edits spread beyond the selection.
These are the failures that show up in real use, with how you'd spot each one.
- Firefly: switching models mid-campaign. A teammate picks a partner model for one image because it handles a scene better. The documented Strength slider applies to Firefly Image models, so that image doesn't follow the dial like the rest, and it now costs credits. Spot it in the model name on the generation; fix it by agreeing one model per campaign.
- Firefly: Strength set too high. At the far right, new images can follow the reference so closely that the whole set starts to look alike. If that happens, back the slider off until scenes vary while the colour and light hold.
- Firefly: the assistant picks the model. Firefly's AI Assistant chooses between Adobe and partner models unless told otherwise. Add "use a Firefly Image model" to the request when the image is going into paid advertising.
- ChatGPT: the long-session wander. Over a long session, colours can drift and faces can change from one image to the next. Start a fresh chat inside the same project for each image, call the saved reference again, and keep a written list of the brand rules in the project instructions rather than relying on the chat's memory.
- ChatGPT: edits that spread. A selection meant for a box label also redraws the hand holding it. Compare the before and after side by side, and reject edits that touched anything you didn't ask about.
- Both: invented marks. Fake logos on vans, plausible-looking number plates and nonsense box labels can turn up in either tool. Zoom to 100% on every image and check the edges of vehicles, clothing and packaging before anything is published.
A before-and-after from the firm shows why the routine matters. Before the trial, its social posts mixed stock photos of red-shirted movers with its own photos of green-shirted crews, so the feed looked like two companies. After: every supporting image came from the same reference at the same Strength, real photos carried the hero spots, and a ten-second zoom check caught the invented marks. Nothing about the new process was complicated; it was just written down.
Can you use Firefly and ChatGPT images commercially, and what happens to your data?
Adobe says outputs from its Firefly models can be used for commercial purposes, and OpenAI's terms say that, as between you and OpenAI, you own ChatGPT output. Brand work also means knowing what you can safely publish. Adobe's plans page says the Firefly models are trained on licensed content and public-domain content where copyright has expired, and that Adobe automatically attaches Content Credentials. For partner models, Adobe says you're responsible for deciding whether the content suits your project, and its IP indemnity is offered only to eligible enterprise customers. Adobe's AI Assistant FAQ adds that your content isn't and won't be used to train generative AI models.
OpenAI's terms say that, as between you and OpenAI, you own the output and OpenAI assigns you its rights "if any", and they warn that output may not be unique: other users may receive similar output. On data, OpenAI's data controls page says sharing for model training is on by default for personal Free, Plus and Pro workspaces; you can turn off Improve the model for everyone under Settings, then Data controls. ChatGPT Business says it doesn't train on your business data by default. If you upload customer photos or unreleased designs as references, settle that setting first. The wider question of using AI images in marketing at all is covered in whether you can legally use AI-generated images in your marketing.
What should an on-brand image control card contain?
A brand-control card for image makers is a one-page table of settings: production model, style reference, Strength, prompt ending, ideas tool, real-photo-only subjects, final check and logo handling. The firm turned its trial into one, kept in the shared marketing folder. Here it is, filled in, so you can copy the structure:
| Rule | The firm's setting |
|---|---|
| Production tool and model | Firefly, standard Firefly Image model, never a partner model for published images |
| Style reference | "Brand-ref-spring.jpg" (crew photo, overcast street) in the shared folder |
| Strength | About 75%; lower to 60% if scenes look repetitive |
| Prompt ending, every time | "no text, no logos, no number plates" |
| Ideas and rough drafts | ChatGPT, inside the "Brand images" project with rules in the instructions; training setting checked |
| Real-photo-only subjects | Vans showing livery, staff faces, customers' homes |
| Final check before publishing | Zoom to 100% on vehicle sides, clothing and boxes; compare colours with the brand swatch |
| Logo and headline | Added in the layout tool from the brand kit, never generated |
A quick sum shows why the card pays for itself. If a twelve-image campaign averages three attempts per keeper, that's 36 generations. In Firefly with a standard model on a paid plan, those cost no credits; switch the same work to Firefly Image 5 at 10 credits a generation and it's 360 of Standard's 2,000 monthly credits. In ChatGPT the generations come out of your plan's limits rather than a visible balance, and OpenAI's pricing page says limits apply without listing an image count, so a long brand session on Plus can run into one. Either way, the card's real saving is in attempts: a fixed reference and a standard prompt ending should cut the number of tries before you get a keeper, and that's worth measuring over your first campaign.
The same split works for a yoga studio, a tutoring centre or a landscaping company: the tool that remembers your look through settings produces the set, the tool you can talk to helps you find it, and a person with a written card checks every image before it goes out. Before publishing, the check in Firefly's style reference and Strength settings is worth running on your first reference image, because a weak reference makes every later image weaker.
If you want help fitting an image workflow like this into how your business actually runs, I offer one-to-one AI consultations for small businesses.
Brand-image questions about Firefly and ChatGPT
Is Adobe Firefly or ChatGPT better for on-brand marketing images?
For repeatable on-brand images, Firefly is easier to control: a style reference with a Strength slider on its own Firefly Image models, model choice, and outputs Adobe says can be used for commercial purposes. ChatGPT Images is easier to steer by conversation, with a saved reference photo and one-area edits. Explore in ChatGPT; produce in Firefly.
Can I use GPT Image inside Adobe Firefly to get both?
Partly. Firefly offers OpenAI's GPT Image models, now led by GPT Image 2, as partner models, and you can start from a reference image there. But Adobe documents the style reference with a Strength slider for its own Firefly Image models, partner models cost credits, and Adobe's commercial-use statement covers only Firefly models.
Should I ask either tool to draw my logo into images?
No. Both tools can produce distorted or invented lettering, and a regenerated logo is not your registered artwork. Generate the scene with no text or logos, then place your real logo file and headline in a layout tool such as Adobe Express or Canva, where they stay exact.
Does ChatGPT train on the brand images I upload?
On personal plans it can, unless you switch it off. On personal Free, Plus and Pro workspaces, OpenAI's data controls page says sharing for model training is on by default; turn off Improve the model for everyone under Settings, Data controls. ChatGPT Business says it doesn't train on business data by default. Adobe says your content isn't used to train its generative models.
Further reads
- Canva AI vs ChatGPT Images: Which Makes Better Social Graphics? — Canva AI against ChatGPT Images for everyday social graphics.
- Adobe Firefly vs Midjourney: Which Is Safer for Client Images? — Which tool is safer when images go to clients.
- How to Upscale a Low-Resolution Logo With Adobe Firefly for Print — Rescue a low-resolution logo or product photo in Firefly.
- Adobe Firefly Commercial Use: What Designers Can Promise Clients — What designers can promise clients about Firefly output.
- How Much Does Ideogram Cost for a Small Business? — What Ideogram Plus, Pro and Team cost, what a credit buys, and how many images each plan makes for a small business.
- Is Adobe Firefly Worth It for a Small Business? Standard vs Pro — A plan-by-plan verdict on Adobe Firefly for a small business, built around what a real month of use costs in credits.
- Canva vs Adobe Express for a Five-Person Team: Brand Locks and Approvals — Canva Business and Adobe Express Teams costed for five seats, with each tool's brand locks, approvals and AI allowance side by side.
Browse all AI-for-business tutorials
Sources: Adobe Firefly help pages on style settings (updated 16 Jun 2026), GPT Image in Firefly (16 Jun 2026), custom models (21 Sep 2026) and AI Assistant brands and FAQ (10 Nov 2026); Adobe Firefly plans page and FAQ; Adobe generative credits FAQ and partner models table; OpenAI help articles Images in ChatGPT, Projects in ChatGPT and Data controls; ChatGPT pricing and Business pricing pages; OpenAI terms of use. All checked 17 Nov 2026.
