How to make AI perfume ads that keep the bottle accurate

To keep the bottle accurate in an AI perfume ad, never describe it in words alone. Give the model approved photos of the real bottle (front, side, top, label close-up), tell it what must not change, generate the scene around it, and inspect cap, label and glass at full resolution before anything goes to the client. Reference images reduce drift but do not guarantee exact reproduction, so a human check or a finishing step stays part of the process.

Why bottles drift in AI images

A text prompt gives the model a category ("a flacon of eau de parfum"), not your product. One guide to Nano Banana 2 notes that a single reference only carries what it shows, and that if the new scene reveals a side the reference never captured, the model has to guess it. For a perfume bottle the guessing shows up in the same places every time: cap geometry, shoulder curve, embossed base, liquid colour and label text.

Small text is the weakest point. In one hands-on test of Flux Kontext, a close-up can ad reproduced the text correctly, but the text failed once the product sat further from the camera (Creative Pad Media). The practical lesson for perfume: a hero close-up of the label is safer than a wide shot where the bottle is small.

A workflow that keeps the bottle intact

1. Build a reference set. Shoot or export clean, neutral-light photos: front, side, back, top view of the cap, and a flat label close-up. If the base is embossed, add it. Runware's guide recommends covering the views your scenes will show, including any side that carries detail the front cannot reveal (Runware).
2. Name what must stay fixed. Write the preservation rule before the creative request: bottle shape, cap, label, logo, proportions and angle unchanged. A product-editing guide from Zeely suggests naming what must stay fixed before asking for the change (Zeely).
3. Generate the still first. Lock the bottle in an image, approve it, then animate. Google's own advice for Veo is to create ingredient images with its image model first, then use them to make videos (Google).
4. Change one thing per pass. Background, then light, then crop. Smaller edits are easier to inspect, which is also the advice in the Zeely guide.
5. Keep the label large in frame. If a logo becomes unreadable when the bottle rotates, one video guide suggests reducing the rotation, showing the label in a separate shot, or using a clean product still at the brand-sensitive moment (Media.io).
6. Inspect at full size. Check cap, label spelling, glass thickness and liquid colour against the physical sample, not against a thumbnail.

Tools and approaches compared

Approach	What it offers for bottle accuracy	Limits to know	Best for
Nano Banana 2 (Google)	Fidelity of up to 14 objects in a single workflow	Fine print can still vary; verify every label	Fast still iterations with several references
Veo 3.1 Ingredients to Video (Google)	Up to three reference images per generation	Not a guarantee of exact reproduction (see Media.io above)	Moving shots once the still is approved
FLUX.1 Kontext (Black Forest Labs)	Edits the scene around an existing product photo	BFL docs now recommend FLUX.2 for new editing projects; small text is fragile	Background and lighting changes on a real packshot
Adobe Firefly Custom Models	Models trained on a specific subject or a visual style	Require an enterprise license and Adobe storage for business	Large brands that want a trained bottle model inside Adobe tools
Miravela	12 image and video models in one place, brand kit applied automatically	Model limits still apply; no guarantee of exact reproduction	Teams that want brief, generation and client approval together

Two points on the table. Training a subject model, as Firefly allows, is a different approach from passing references at generation time, and it can pay off when the same bottle appears in hundreds of assets. The cost is that the model lives inside one vendor's ecosystem. Reference-based generation with general models is lighter to start with and works across vendors, but you carry the verification burden yourself.

Where Miravela fits, and where it does not

Miravela is an image and video production platform for brand teams and agencies. It currently offers 12 hand-picked models, including Nano Banana 2, GPT Image 2, Flux Pro, Veo 3.1 and Kling, and it picks one per job. Your brand kit (logos, colours, references, rules) is applied automatically, and clients review and approve in a private portal where comments stay inside the platform. For perfume, that matters in two ways: you can use a precise still model for the bottle and a video model for motion without switching subscriptions, and the client signs off on the bottle in the same place where it was made.

Miravela does not claim to guarantee that a generated bottle is pixel-identical. The models run at their providers, and the limits described above come from those models. If your priority is a custom-trained bottle model inside Photoshop, Adobe's route is the better fit. If your priority is retouching a real packshot photo you already own, a dedicated editing workflow with a human retoucher on the label is safer than any generator.

What Miravela platform data says about iteration

These figures come from Miravela's own database over the last 90 days, aggregated, with no client named. GPT Image 2 ran 3,074 generations with a 95.9% completion rate and a median time of 95 seconds. Nano Banana 2 ran 799 generations with an 89.5% completion rate and a median of 26 seconds.

For a bottle workflow, this means Nano Banana 2 gives you roughly three to four attempts in the time one GPT Image 2 run takes, so it suits exploring compositions and angles. The lower completion rate means about one in ten started runs does not finish, so plan for a retry. A slower model can be kept for the final hero. One third-party guide makes a similar split, suggesting GPT Image 2 when on-image text must stay accurate (Gen NanoBanana); test that on your own label rather than assuming it.

Video is slower. In the same period, Kling 3.0 had a median time of 197 seconds and Seedance 2.0 a median of 312 seconds. Each bottle check on a clip therefore costs minutes, which is why locking the still first saves time.

How to choose

• Do you have clean photos of the real bottle from several angles? If not, shoot them first; no tool compensates.
• Is the label text small in the final frame? If so, crop closer or plan a compositing step for the label.
• Will the same bottle appear in very large volumes? Consider trained subject models; otherwise references are enough.
• Do you need video? Approve the still before you animate, and check the label on the first and last frames.
• Who signs off? Choose a setup where the client sees the selection and comments in one place, not in email threads.
• Can you export what was asked of the model? Miravela keeps the prompt history of every visual and makes it exportable, which helps if a brand must document its creative input. Plans start at €199/month excl. VAT on the pricing page.

For more on the model line-up, see the models page; agencies running several fragrance clients can look at the agency page.