Photograph the product badly on your phone, then describe the ad you want in one plain sentence. The agent picks the model and writes the prompt. Two finished ads from one snapshot cost 23 cents here — and the wordmark survived intact while the label's die-cut shape quietly changed.
Product photography is a solved problem if you have a studio, a light kit and a day. For everyone selling something out of a spare room, the actual constraint is that a phone snapshot on a kitchen counter does not look like something anyone wants to buy.
This is the workflow for closing that gap without learning photography or prompt engineering, tested on a deliberately bad photograph — plus an honest account of what the model quietly changes while you are not looking.
Want to try it on your own product? Rangy's agent takes a photo and a sentence, and picks the model for you.
Try it freeHow does the agent workflow actually work?
You import a photo, select it, and type what you want in ordinary language. There is no model picker, no prompt template and no settings to learn — the agent chooses which model suits the job, writes the detailed prompt itself, and runs the generation, narrating what it did and what it cost.
That is the genuine difference from the guides on this site that teach you to choose a model and write a good prompt. Those skills produce better results and they are worth having. The agent route trades some of that control for not needing them at all.
Underneath, it is a planner with a tool set rather than a chat wrapper. It can generate images and video, upscale, save and recall brand assets, keep per-project memory, and ask you a clarifying question when the brief is ambiguous. If a model fails mid-run it falls back to a related one rather than stopping.
The prompt that produced the studio image above was one sentence: create a professional advertising image of this product. Everything else — the seamless backdrop, the softbox highlights down the glass, the reflection under the jar, removing the plug socket and the bread bag — came from the model's own reading of "professional advertising image".
What makes a good source photo?
Sharpness, not styling. The lighting, the background and the angle are all things the model will replace anyway. What it cannot replace is detail that was never captured — if your label is blurry in the original, it will be wrong in the output.
The snapshot used here was deliberately poor: flat ceiling light, a yellow-green cast, visible clutter, shot off-level from too close. All of that was fixed. What survived cleanly was the label text, because it was legible in the source.
- Get the product sharp. Wipe the lens, brace the phone, tap to focus on the label. This is the only part that matters.
- Fill more of the frame than feels natural. Detail you did not capture cannot be recovered.
- Ignore the background. Clutter is genuinely fine; it gets replaced.
- Ignore the lighting. Also replaced. Bad light is not the problem, blur is.
- Shoot the label straight on if the text matters, so the model has an undistorted reading of it.
Does your actual product survive?
Mostly, and the exceptions are specific. Across both generations the wordmark, its spelling, its typeface and its position on the jar came through identically. What changed was the label's die-cut shape — a plain rectangle in the original became notched corners in one output and bowed edges in the other.
That is the pattern worth internalising, and it is the same one measured in the garment fidelity test: what a customer reads survives, and what a customer could measure drifts.
- Reliable: brand name, label text, colour, overall silhouette, material impression.
- Unreliable: die-cut shapes, exact proportions, cap threading, embossing, small print, seams.
- Different every run: the drift is not consistent, so two ads of the same product can disagree with each other.
The practical rule: use generated images where the job is to make someone want the product, and a real photograph where the job is to show them precisely what arrives. For most sellers that means lifestyle and ad creative generated, and the main listing image shot.
How do you get the variant you want?
Ask for it in plain words and let it re-run. "Make it vertical", "make the background white", "give me one for Instagram stories" all work as follow-up messages, in any language, without restating the original brief. The agent keeps the context of what it just made.
This is where the agent route earns its keep. Iterating by conversation is much faster than editing a long prompt, because you are describing the change rather than rewriting the whole specification each time.
Three practical habits:
- Change one thing per message. "Make it vertical and brighter and add props" produces something you cannot diagnose.
- Ask for several at once when exploring. A run of variations costs cents and the best of three beats the first of one.
- Name the platform, not the numbers. "For Instagram stories" is understood; you do not need to know the pixel dimensions.
How many images does one product need?
More than most small sellers make. A listing typically wants a clean pack shot, a scale reference, a lifestyle scene and a detail crop, and a product with any following also needs social and ad creative. That is five or six images per item, which is exactly why the per-image cost matters.
The reason this used to be a problem is that each extra image was another setup, and a shoot day only produces so many. When a scene costs cents, the constraint becomes deciding what to show rather than affording to show it.
- A clean pack shot on plain white, for the listing. Shoot this one if you can.
- A scale reference next to something familiar, which prevents a common complaint.
- A lifestyle scene showing the product in use or in context.
- A detail crop of the material or finish, which is where a real photograph still wins.
- Format variants — square, vertical, and whatever your ad platform wants.
What does it cost per image?
The two ads at the top cost nine cents each, on top of five cents for the source snapshot — 23 cents for the set. A real product shoot for one item runs into the hundreds once a photographer, a studio and styling are counted, and takes days rather than minutes.
| Step | What it does | Cost |
|---|---|---|
| Reference-guided edit | Turns your photo into an ad, keeps the product | ~$0.09 at 2K |
| A cheaper edit model | Fine for background swaps and simple scenes | ~$0.06 at 2K |
| Three variations | Explore angles and settings in one go | ~$0.27 |
| Upscale a keeper | Take the winner to print or listing size | ~$0.05 |
Rates from Rangy's live pricing tables at the cheapest configured provider, checked 15 August 2026. You pay the provider directly on your own key; Rangy does not bill for generation.
The number that matters is not the saving on one image. It is that trying eight scenes for a product becomes a 70-cent decision rather than a second shoot, which changes how much you are willing to experiment.
Will marketplaces accept these images?
Policies differ by platform and have been revised repeatedly, and several now ask you to declare AI-generated imagery. The safer principle than any current rule: the image must not misrepresent what the buyer receives, which is a standard that predates AI and applies regardless of how the picture was made.
Because the drift documented above is real, the honest reading is that generated images are well suited to lifestyle shots, ad creative and social content, and poorly suited to the main listing image where a buyer is judging exactly what they are buying.
Check the current policy of the platform you sell on rather than relying on an article. The marketplace-specific detail is covered in the e-commerce product photos guide.
Where does this approach fall down?
Blurry source photos, products with fine print or intricate hardware, anything where exact proportions matter, and cases where you need every image in a range to agree with the others. The agent removes the learning curve; it does not remove the underlying model's limits.
Stated plainly, because this guide is published by the company that makes the tool:
- Detail that was not in the photo cannot appear. The video demonstrates this directly — a small area that was unclear in the original stayed wrong in the output, and the narration says so.
- Ingredient lists, warnings and legal copy will not survive intact. Composite real label crops back on, or shoot those images.
- Consistency across a range takes work, because the drift differs per run. Re-anchor every generation to the same source photo rather than chaining outputs.
- The agent is still early access at the time of writing, and the video says the same — it is being developed actively rather than finished.
- If the product is the craft — jewellery, ceramics, anything where texture is the selling point — a real photograph earns its cost.
Frequently asked questions
Can AI turn a phone photo into a professional product image?
Yes. A deliberately poor snapshot — flat ceiling light, colour cast, visible kitchen clutter, shot off-level — became a clean studio advertising image and a styled lifestyle scene for nine cents each. The lighting, background and framing are all replaced. What cannot be replaced is detail that was never sharp in the original.
Does the AI keep my actual product the same?
The parts a customer reads survive well. In this test the brand name, its typeface and its position were identical across both outputs. The label's die-cut shape changed though, from a plain rectangle to notched corners in one image and bowed edges in the other. Text and colour are reliable; exact shapes and proportions are not.
What kind of photo should I take first?
A sharp one. Ignore the background and the lighting, because both get replaced anyway. Wipe the lens, brace the phone, tap to focus on the label, fill more of the frame than feels natural, and shoot the label straight on if the text matters. Blur is the only fault that carries through to the result.
Do I need to know how to write prompts?
Not for this route. The agent reads a plain sentence, chooses which model suits the job and writes the detailed prompt itself. Learning to prompt and pick models still produces better and more controllable results, which is what the other guides on this site cover, but it is no longer the entry requirement.
How much does an AI product photo cost?
About nine cents for a reference-guided edit at 2K on your own provider account, or around six cents on a cheaper edit model. Three variations run about 27 cents. You pay the model provider directly rather than a subscription, so a month with no work costs nothing.
Can I ask for different sizes and formats?
Yes, in plain language and as a follow-up rather than a new brief. Asking to make it vertical, change the background to white, or produce one sized for Instagram stories all work, and you do not need to know the pixel dimensions. Prompts in languages other than English are understood as well.
Are AI product photos allowed on marketplaces?
Policies differ by platform and change, and several now require you to declare AI-generated imagery. The durable principle is that an image must not misrepresent what the buyer receives. Given that fine shapes drift between generations, generated images suit lifestyle and ad creative better than the main listing image.
The bottom line
Take one sharp photo, describe the ad in a sentence, and iterate by conversation. Use the results for lifestyle and advertising, keep a real photograph as the main listing image, and check the label at full size on every keeper before it goes anywhere.
The part that genuinely changed is the entry cost. Turning a bad snapshot into something that looks like an ad used to require either money or a skill; it now requires a sentence and about a quarter.
What has not changed is that the picture still has to be honest about the thing in the box. That is the one judgement no model makes for you.
"This software has increased my workflow speed tenfold, and the output quality it has delivered in my work has been exceptional."
One photo, one sentence
Rangy's agent takes your product photo and a plain-English brief, picks the model, writes the prompt and shows you the cost before it spends anything — with every image saved to your own disk.
Download Rangy Free Or watch the full walkthrough →Mac & Windows · Free plan, no credit card · 5 generations a day
The three images at the top are one real run, shown unretouched. The source snapshot was generated with GPT Image 2 at 2K for $0.05, deliberately specified as a poor phone photograph with clutter, a colour cast and an off-level angle, so that it would stand in for a genuinely bad product shot. Both advertising images were produced with Nano Banana Pro at 2K for $0.09 each, with that snapshot attached as a reference image and an instruction not to redesign the product or its label. The label comparison is a 100% crop of the same region in all three files; the finding that the wordmark held while the die-cut shape changed comes from inspecting those crops, not from an assumption. Model rates come from Rangy's live pricing tables, checked on 15 August 2026, and Rangy does not bill for generation — you pay the provider on your own key. No marketplace's current AI-disclosure policy is quoted, because they differ and have been revised repeatedly.
One caveat stated plainly: the source photograph is itself generated, because no real product was available to shoot. That makes this a fair test of what the edit step preserves and changes, and not a test of how a real camera file behaves. The agent is described from its documented tool surface and from the linked walkthrough, and it remains in early access at the time of writing.
This guide is published by Rangy, which makes the tool it describes, so it is not a neutral source. The case against using these images for your main listing is in does your product survive and where this falls down.