One good product photo, and a campaign that calls for three different moods. Anyone working in product marketing knows the situation. Every extra shoot costs time and money. So I wanted to know: how far can you get with one existing photo and an image model that only changes the surroundings?
I didn’t leave it at a question. In a hands-on test, we built a small pipeline in our own XS2Content platform and ran it three times on the same sneaker photo. Below I show how that worked, what came out of it and where the limits are.
The starting point: one photo, one question
The original: a colourful pair of sneakers on a light blue background, with coloured blocks, transparent discs and a grey pedestal. A fine photo, with a very distinctive mood. The question: can we leave the shoes as they are and replace only the setting, without a photographer or retoucher?
How the pipeline is put together
Deliberately kept small, in two steps:
- Input (FormInput). Exactly two things go in: the product photo and a scene prompt – a plain-language description of the desired surroundings.
- Image editing with FLUX.2 Pro. Photo and prompt go to an editing step based on FLUX.2 Pro by Black Forest Labs, via Replicate. Set to 1 megapixel, 16:9, png, one image per run.
Then comes the most important step, and it isn’t automatic: assessing the result. With my own eyes.
This was a standalone test pipeline: not scheduled and not connected to the website, social media or any other channels. Every prompt started with the same instruction: change only the surroundings and leave both shoes exactly as they are – same position, perspective, colours, materials, laces and soles. No text, no extra logos, no extra objects.
One choice matters: for all three scenes we always used the original photo as input, never a previously generated image. That way you don’t stack deviations on top of each other. Each scene got one run, with no second chances and no tweaking in between. What you see below is therefore the first attempt.
Scene 1: neutral studio
Requested: a premium, neutral studio in warm light grey, with soft contact shadows.
The busy background is gone and the shoes stand calmly in the frame. Only the studio has turned taupe, clearly darker than requested. Usable as a mood image, not exactly what the prompt described.
Scene 2: a New York street
Requested: the shoes on a New York sidewalk, photographed low to the ground, in daylight, with a strongly blurred street with buildings and yellow taxis.
Buildings, blur and yellow taxis make the scene instantly recognisable. Depth, light and shadow feel like a real street photo. But here too, the model doesn’t leave the shoes untouched: perspective and light on the shoes shift along with the new surroundings.
Scene 3: from the stage into the auditorium
Requested: the shoes on a wooden stage, with the camera pointing from the stage towards an empty auditorium with red seats and golden balconies.
The mood is there instantly: warm, festive, a touch theatrical. Look closely and you’ll see a small dark object behind the shoes that wasn’t in the brief. Exactly the kind of detail you catch in a review.
What this does and doesn’t show
To be fair: the product stays recognisable, but it isn’t identical. Colour blocks and shape match the original. Look at the details, though, and material, laces, soles, perspective and lighting have changed in all three images. The output is also slightly smaller than the original (1344×752 versus 1360×768).
For mood images, moodboards and quick campaign explorations, I already find this interesting: you can visualise an idea without a new shoot. But where the product has to be exactly right – a webshop photo, a specific variant, anything a customer bases a purchase on – this isn’t enough. This setup doesn’t guarantee pixel-perfect product preservation. That requires a different approach, for example with masks that protect the product and a check afterwards that measures deviations.
Two lessons I’m taking with me:
- A prompt is not a contract. The model follows the direction, not every word.
- Human review is not a formality. It’s the moment you decide whether an image is ready to go out.
What this means for us
This is how we approach new AI capabilities at XS2Content: not claiming that something “can do everything”, but building small, actually running it and looking honestly at what comes out. This test showed that an existing image step in our platform can create strong scenes – and where the limit lies.
So the question isn’t whether this works, but for which kinds of images.
Do you work with product photos and want to know where this does and doesn’t fit in your campaigns? Send me a message – I’m happy to think along with you.
— Sebastian Plasschaert, XS2Content
