Most AI product photography tools do one thing. You upload a photo, they cut out the background, and you get a nicer version of the same single image back. Then you still have to write the caption, still have to stage a lifestyle shot, still have no video, and still have nothing to post tomorrow.
The interesting question is not how to improve one photo. It is how to get an entire week of content out of one photo.
This is a walkthrough of a pipeline that does exactly that. One phone snapshot goes in. Six finished assets come out: a clean studio pack shot, three lifestyle scenes, a product video and a full posting kit. Roughly 54 credits and under ten minutes, most of it spent waiting on the video.
Everything below is from a real run, including the three things that broke.
The input is deliberately bad
A hand poured soy candle in an amber glass jar, sitting on a cluttered kitchen table. Mixed window and ceiling light. A used mug, a folded tea towel, a set of keys and some crumbs in frame. Shot from standing height, slightly off centre, no filter.
That is not a mistake. That is the point. It is what a real seller actually has on their phone, and if a pipeline only works on photos that were already good, it does not solve anything.
What your input photo genuinely needs
- One product, nothing overlapping it
- Daylight if possible, any light if not
- The label facing the camera and readable
- JPG or PNG, because the video step rejects webp
A messy background is fine. Bad exposure is fine. Your phone is fine.
The one idea that makes the whole thing work
Here is the part most people get wrong when they chain AI steps together.
The naive approach is to feed your original photo into every step. Generate a lifestyle shot from it, generate a video from it, generate a flat lay from it. The result is four images of four subtly different products. The label drifts. The proportions shift. The lid changes shape. Nothing matches, and a customer scrolling your grid can tell, even if they cannot say why.
The fix is to make one image the anchor.
Step one turns the phone photo into a clean studio pack shot. You look at it, and you check it against the real product in your hand: the label wording, the logo, the silhouette, the proportions. If it is wrong, you fix it and run again. It costs 7 credits to reach that checkpoint.
Then every step after that reads from the pack shot, not from your original photo. The scenes, the video, the copy. All of them.
The whole trick in one line
That single design choice is why the jar in the flat lay is the same jar as the one in the video. And it is why one approval gate protects every credit spent after it.
What actually came out
The studio pack shot
Same amber glass, same cream wax line, same single cotton wick, same kraft label with the brand name in widely spaced serif capitals. Set on a seamless warm off white backdrop with one soft light source and a grounded contact shadow.
Nothing was invented and nothing was lost. That is the bar for ecommerce product photography: the picture has to match what arrives in the box.
Three lifestyle scenes
Rather than one generic "put it somewhere nice" prompt, the pipeline runs three separate directors, each with a fixed job.
Splitting the job into three fixed roles is what makes this lifestyle product photography rather than three variations of the same shot. It also means you can rerun one scene without touching the other two.
One detail worth noticing: the flat lay chose figs on its own, because the brief said the scent was fig and cedar. It read the brief rather than decorating at random.
Turn a photo into a video with AI
The pack shot goes into the video model as the first frame, which means the clip literally opens on the image you already approved and turns your actual jar rather than something that merely resembles it. Ours came out 960 by 960 at just over five seconds, rotating roughly a quarter turF
That last detail matters, and it is the honest limit on 360 product photography done this way. Whether you get a quarter turn or a full loop depends entirely on which video model you pick, and there is a section on that below.
The posting kit
Caption, hook, three alternate hooks, ten hashtags and alt text. Written from the brief and the pack shot, in the tone and for the platform you specified.
Ready made swaps: make it yours in five minutes
The canvas is a template, not a candle advert. Here is exactly what to change and the text to paste in.
Two rules before you start:
1. Instructions go in SYSTEM PROMPT, not USER PROMPT. When a Prompt node is wired into an LLM node, that connection replaces whatever is typed in the user prompt field. Anything you type there is silently ignored.
2. Changed your photo or your brief? Rerun the pack shot director and the pack shot first. Everything downstream reads from the pack shot, so a stale pack shot quietly poisons all five other assets.
Swap 1: your product
Drop your photo in the IMAGE node, then paste this into the PRODUCT BRIEF node and fill it in.
PRODUCT BRIEF:
Product: [what it is, in five words]
Brand: [brand name]
What it is: [one sentence a stranger would understand]
Materials and finish: [glass, brushed steel, matte black metal, kraft label, and so on]
Who buys it: [age, situation, why they care]
Where they use it: [three real places]
Tone: [calm and premium / loud and fun / clinical and precise]
Platform: [Instagram / TikTok / Amazon listing]
CTA: [shop the link in bio / order on WhatsApp / free shipping this week]Materials and finish and Who buys it are the two fields that change the output most. Leave them vague and every scene comes back generic. Write "brushed stainless with a knurled grip, bought by people who cook every night" and the scenes change completely.
Swap 2: the backdrop of your pack shot
Open the PACK SHOT DIRECTOR node, find this line in the system prompt, and replace it:
place the product on a seamless studio backdrop in a soft neutral tone that flatters its own colour| You want | Paste this instead |
|---|---|
| Amazon and marketplace compliant | place the product on a pure white seamless background, RGB 255 255 255, with a soft contact shadow only and no other shadow or reflection |
| Warm and premium (default) | place the product on a seamless studio backdrop in a soft warm neutral tone that flatters its own colour |
| Stone or marble slab | place the product on a honed pale marble slab against a soft grey seamless backdrop, with a faint natural reflection on the stone |
| Brand colour gradient | place the product on a smooth vertical gradient backdrop running from [YOUR HEX] at the top to a lighter tint of the same hue at the bottom |
| Dark and moody | place the product on a deep charcoal seamless backdrop lit from behind and one side, with a bright rim highlight along the edge of the product and the background falling into near black |
Want a video hook that stops the scroll without reshooting a single frame?
Learn how one ordinary cooking clip got a single impossible moment; a mid-pan flambé that vanishes as fast as it appears, using Seedance 2.0 Video Edit in Zoviz Pro Studio.
See the full hook trick tutorial →
Swap 3: the three scenes for your category
The three scene directors are set to in use, at rest and flat lay. Keep the structure and change the world.
| Category | In use | At rest | Flat lay |
|---|---|---|---|
| Skincare | applying it at a bathroom mirror in morning light | on a bathroom ledge beside a folded towel and a small plant | with a linen cloth, a ceramic dish, and two or three raw ingredients from the formula |
| Food and drink | pouring or serving it at a kitchen counter mid morning | on an open shelf beside jars and a wooden board | with the raw ingredients, a linen napkin and a fork or spoon |
| Apparel | worn mid stride outdoors in overcast daylight | folded on a chair beside a bag and a pair of shoes | flat on a plain surface with the folded garment, a belt and shoes at the edges |
| Tools and hardware | held mid task on a workbench under a task lamp | hung on a pegboard beside other well kept tools | with the tool, its bits or accessories, and the material it cuts or fixes |
| Jewellery | being fastened at the wrist or neck in soft window light | on a dish beside a folded silk and a candle | on textured paper with dried flowers and a velvet pouch |
Swap 4: the final video output
This is the one everybody wants to change, and it is the easiest. Replace the whole contents of the SPIN DIRECTION prompt node with one of these. Each one keeps the pack shot as the first frame, so the clip always opens on your approved image.
Turntable loop:
A slow controlled product turntable. Start exactly on the first frame, unchanged. The product rotates smoothly on the spot through a full 360 degrees at a calm even speed and returns to the exact starting position on the final frame, so the clip loops seamlessly with no visible cut. The camera is locked in place at product height, no zoom, no push in, no shake, no cuts. The backdrop, the lighting and the exposure stay identical for the entire clip, and the contact shadow turns naturally with the product without detaching from it. The label and the logo stay sharp and readable through the whole rotation and never warp or change their wording. Soft ambient room tone only, no music, no speech, no on screen text.Hero push in:
Start exactly on the first frame, unchanged. The camera pushes in slowly and smoothly toward the product on a locked axis, ending on a tight framing of the label with the product still perfectly upright and centred. The move is continuous and even with no acceleration, no shake and no cuts. The lighting and backdrop stay identical throughout while the depth of field softens slightly as the camera closes in. The label and the logo stay sharp, readable and unchanged. Soft ambient room tone only, no music, no speech, no on screen text.Hands reaching in:
Start exactly on the first frame, unchanged. A single pair of real hands enters the frame from the lower right, picks the product up gently, turns it once so the label faces the camera, and holds it steady. The camera stays locked in place. Movement is slow, natural and unhurried. The product stays identical throughout with its label sharp and readable, and the hands never obscure the brand name. Lighting and backdrop remain constant. Soft ambient room tone only, no music, no speech, no on screen text, no distorted or extra fingers.Light sweep:
Start exactly on the first frame, unchanged. The product stays completely still and perfectly centred while a single soft highlight travels slowly across its surface from left to right, revealing the material and finish as it passes. The camera is locked with no movement at all. Nothing about the product changes: same shape, same colour, same label, same position. The backdrop stays constant while only the highlight moves. The move is slow, continuous and even. Soft ambient room tone only, no music, no speech, no on screen text.Texture and steam:
Start exactly on the first frame, unchanged. The product stays still and centred while soft steam rises gently and irregularly from it, drifting upward and dissipating naturally. The camera is locked in place with no movement. The product itself does not move, rotate or change: same shape, same colour, same label, same position. Lighting and backdrop stay identical. The motion is subtle and calm rather than dramatic. Soft ambient room tone only, no music, no speech, no on screen text.Fabric drape:
Start exactly on the first frame, unchanged. The fabric settles and shifts very slightly as if caught by a faint breeze, with the folds moving softly and naturally while the garment stays in the same position and framing. The camera is locked with no movement. Colour, texture, print and any visible label stay identical throughout. The motion is gentle and continuous with no cuts. Soft ambient room tone only, no music, no speech, no on screen text.Swap 5: platform and tone
You do not need to touch the copy node at all. Change Platform: and Tone: in the product brief and the caption, the hook and the hashtags all follow. Set Platform: Amazon listing and you get listing copy. Set Platform: TikTok and Tone: fast and funny and you get something entirely different from the same photo.
Using this for AI UGC ads
The scene director set up for AI UGC ads is slightly different from the ecommerce one, and it is worth calling out because the two get confused.
Marketplace photography wants the product isolated, evenly lit and unambiguous. UGC wants the opposite. It wants the shot to look like somebody's camera roll: a bit of clutter, a real hand, light that came from a window rather than a softbox, a phone framing that is slightly off.
Three changes turn this pipeline into a UGC generator:
- In the brief, set Tone: to something conversational and name the platform as TikTok or Reels
- In the in use scene director, add shot on a phone at arm's length, slightly imperfect framing, ordinary domestic lighting, mildly cluttered background to the end of the description block
- Use the hands reaching in video prompt from Swap 4 rather than the turntable
Keep the pack shot anchor either way. Even a scruffy UGC frame should show the product you actually ship, or you are back to a refund problem rather than a creative one.
No camera, no crew, no actor, one full short film in a single day.
See how a three minute episode called 06:00, complete with a consistent AI lead, fourteen scripted shots, and diegetic sound, was built entirely from prompts in Zoviz Studio.
Read the full prompt-by-prompt recipe →
Three things that broke, and how to avoid them
Every walkthrough you read online works perfectly. This one did not, and the failures are more useful than the successes.
1. The instructions were silently ignored
The first run of the pack shot director returned Instagram captions instead of an image prompt. The cause: a Prompt node wired into an LLM node's prompt input replaces the typed user prompt rather than joining it. The model only ever saw the brief, with no instruction attached, so it did the most obvious thing with a product description.
There is no visual signal that this has happened. The output is plausible, just the wrong kind. Put instructions in the system prompt.
2. Firing three image generations at once hit a rate limit
Two of the three scene images failed immediately with a 429, resource exhausted, from the image model. They succeeded on retry and the failures did not charge, but if you are watching this happen you will assume the pipeline is broken. Run them one at a time, or just retry.
3. Changing the video model deleted a connection
The pack shot was wired into the video node's first_frame port. Switching to a different video model, which has no first_frame port, deleted that connection without warning. The node quietly fell back to text to video and generated a video from the spin prompt alone. The prompt says "start exactly on the first frame, the studio pack shot", but no image was ever attached, so the model invented a product from a description that assumed one existed.
Check this before you spend credits on a video
The result looked fine and had nothing to do with the actual product. Make sure the node header says first frame to video and not text to video.
The bonus fourth thing: your seamless loop may not be seamless
This is worth knowing before you pick a model. Writing "returns to the exact starting position so the clip loops" into the prompt does nothing on a model that only accepts a first frame. It has no mechanism to end anywhere in particular.
A true seamless loop needs a model that accepts both a first frame and an end frame. Feed it the same pack shot on both ports and the clip is forced to start and end on the identical image. First frame only models are cheaper and produce a perfectly good clip, it just visibly jumps on repeat. The loop is a model feature, not a prompt feature.
What it costs
| Step | Credits |
|---|---|
| Pack shot director | 1 |
| Studio pack shot | 6 |
| Three scene directors | 3 |
| Three lifestyle scenes | 18 |
| Copy kit | 1 |
| Subtotal, four images and the copy | 29 |
| Video, first frame only model | 25 |
| Video, first and end frame model with a true loop | 40 |
| Full kit | 54 to 69 |
The ordering matters as much as the total. The expensive step runs last, off an image you have already approved. You reach the decision point for 7 credits.
Is AI product photography actually allowed?
Yes, with one hard limit that is worth stating plainly.
The pack shot must be an honest representation of the product you ship. Same size, same colour, same finish, same contents. Marketplaces reject listings whose main image misrepresents the item, and customers return things that do not match the picture, which costs far more than a photographer would have.
The lifestyle scenes are staged visualisations, exactly as a photographed lifestyle shot is staged. Style them freely. Just never show an accessory, a size, a colour variant or an included item that does not actually ship in the box.
That is not a legal disclaimer, it is the reason the pipeline has an approval gate after step one.
FAQ
What is the best AI product photo generator?
For a single background swap, almost any of them will do and they mostly produce similar output. The question is worth reframing: most tools generate one image at a time with no memory between generations, so your fifth image does not match your first. If you need more than one asset per product, choose a system that lets one approved image anchor everything downstream. That consistency, not the quality of any single render, is what separates a usable set from a folder of near misses.
Can I turn a photo into a video with AI?
Yes. Feed your approved pack shot into a video model as the first frame and it becomes the opening frame of the clip. That is the difference between a video of your product and a video of something that looks a bit like your product.
How many photos do I need to start?
One. That is the entire premise. A better photo produces a better pack shot, but one honest phone snapshot is the requirement.
Will the product look the same across all the images?
It will if every step reads from one approved anchor image instead of from your original photo. If a workflow generates each asset independently from the source, expect drift, and check the label wording in each output before you post.
How long does a full run take?
Under ten minutes for six assets, most of which is waiting on the video.
