Plien Roob
Start freeOpen studio

Research notes

How this is built

Anyone can buy the same image models we use. What separates the results is the instruction sent to them — this page explains how ours is assembled, and which broken image each rule came from.

Countless test runs went into finding the right mix — so that it stays simple to use for everyone.

  • 65

    instruction blocks across the system

  • 360

    scenes labelled by hand for framing

  • 0%

    wrong-shape images after the fix, down from 48%

  • 9.8M+

    person × pose × scene combinations

The uploaded product photo: a long sleeveless cream knit dress with gold buttons all the way down, a belt with a gold spiral buckle and two flap pockets, on a beige background with no one in the shotUploaded
The result: a model with a pink bob and glasses, wearing the same cream dress, walking in front of the Arc de Triomphe in ParisResult
  • A model with a pink bob and glasses, wearing the same dress, leading a camel across the dunes at sunset
  • The same model walking down pavement steps beside an old stone building, with a wooden bench alongside
  • The same model seated in an upholstered armchair in a cream living room, with a gold vase behind the chair
  • The same model standing in a hall of marble columns, with sunlight coming through tall windows
  • The same model sitting on a giant knight piece on a chessboard floor under an open sky
  • The result: a ghost mannequin image of the same dress, shaped in three dimensions and turned slightly, on a white background with no person or mannequin in the shot
The first image is the uploaded product photo. The other 7 came from one press of Image set: the 6 model shots each got a different scene with the pose following it, and the last one is a ghost mannequin — every one of them got its own freshly assembled instruction.

Our instruction isn't one block of text — it's rebuilt for every image

There is no prompt box anywhere in the product. You pick from a library that pairs person × pose × scene in more than 9.8 million combinations, and the system writes the instruction.

Across the whole system, instructions are assembled from 65 blocks, each responsible for something different: Product to Model 25 · Ghost Mannequin 11 · Product Showcase 10 · Product-Sexy 2, plus 17 shared by several tools (casting · poses · styling). Each image draws on them differently, based on what you picked.

What those blocks enforce

  1. Fabric and colour must match the real item
  2. Light on the person must come from the same direction as the scene
  3. The product must not be hidden by hair or hands
  4. Its size against the body must be plausible
  5. The camera must match the scene
  6. Never return a multi-panel collage

Every instruction is reproducible. Nothing is randomised in a way we can't replay — when you say “this one came out wrong”, we can rebuild that exact instruction in full.

Every rule came from an image that actually failed

Everything in this section was distilled from more than 700 images we had AI generate, in the live system and in tests.

48%

Before the fix · 25 real jobs

0%

After the fix · 57 jobs re-measured

The “automatic” aspect-ratio setting used to mean sending nothing at all and letting the vendor guess. Measured across 25 real jobs: 48% came back the wrong shape, the worst one drawing the product a second time to fill the empty space. After the fix, measured across 57 jobs: 0%.

We once ran ten different product types back to back without picking a model, and got nearly the same face every time — even though the instruction asked for variety.

One model invented a fake brand label at the collar despite an explicit instruction not to. We stopped using it for the tools that must stay faithful to the product.

Some scenes let a person stand full-height but render them so small the product is unreadable. So all 360 scenes were labelled by hand for which framing each one can carry.

  • The red Giant Swing at nightFull bodyHalf body
  • A white European-style room with a carved upholstered armchairFull body
  • A Chinatown street lined with shop signsHalf body
  • A stage with red velvet curtains and a heart-shaped cut-out boardFull bodyHalf body
  • A decorated round frame looking out over a rock town on a hillHalf body
  • A desert at sunset with a camel standing on the dunesFull body
  • A golden carousel lit up at nightHalf body
  • A stone terrace by a lake with snowy mountains behindFull bodyHalf body
A few scenes from the library, each with the framing label it was given by hand

Faces that don't repeat — and why asking the AI for variety doesn't work

Each image is generated independently. The vendor has no memory of who it drew a moment ago, so writing “vary the casting” into a single instruction barely moves it — it still lands on the most likely answer every time.

So the variety has to come from our side: we pick the look first, then write that look into the instruction. The model gets a genuinely different brief for each image.

  • Asia50
  • Europe35
  • Africa15

The 10 looks the system draws from — each segment's length is the chance that look gets picked

The current spread is Asia 50 · Europe 35 · Africa 15, across 10 looks. That ratio is the owner's decision based on the market we sell to, not an arbitrary default.

There is no skin-tone instruction anywhere in this system, deliberately — a leftover skin-tone line was exactly what made non-Asian looks impossible in the first place.

Automated tests hold both the declared ratio and the actual distribution over thousands of simulated runs.

As shown, verifiable — a QR code that lets buyers compare the AI image with the real item

A beautiful AI image makes buyers wonder whether the real item will look the same. Our answer is not to hide that AI was used, but to let buyers check for themselves in a few seconds.

The shop switches it on under a result in the studio and gets a check page for that image. The page puts the product photo the shop uploaded next to the image in the post, and says plainly which parts are real and which AI made — the fabric pattern, colour and lettering are real; the model, pose, scene and light are AI.

Shop-confirmed · Compare with the real photo
The uploaded product photo: a long sleeveless cream knit dress with gold buttons all the way down, a belt with a gold spiral buckle and two flap pockets, on a beige background with no one in the shotReal item
The result: a model with a pink bob and glasses, wearing the same cream dress, walking in front of the Arc de Triomphe in ParisImage in the post

As shownVerifiableAI modelScan for the real photo

Confirmed by the shop

  • Fabric pattern
  • Colour
  • Lettering on the item

Created by AI

  • The model (not a real person)
  • The pose
  • The scene and location
  • Light and shadow
A mock-up of the check page, made from the same dress as the top of this page · the QR code in the mock-up leads back here

The downloaded image carries a small QR card in a corner the shop chooses. Buyers scan it and the check page opens straight away. The code spans 14% of the image's short side; we measured that it still scans after the image is shrunk to 40% when posted.

The label speaks for the shop, not for us — the shop ticks to confirm every time. The wording changes per tool: a ghost mannequin is never called an AI model, and a try-on made with the shop's own person photo says the person was not created by AI.

It works with every tool that builds an image from the shop's product — Product to Model (including same-model and front + back), Virtual Try-On, Product Showcase, Product-Sexy and Ghost Mannequin. It costs no credits, the image is assembled on the shop's own device, and it suits posts on IG, Facebook, LINE and the shop's website.

How we choose a model, and why others were cut

The seven tools in the next section do not share one AI. Each was chosen by running the same product photo through several and comparing the results side by side, rather than trusting marketing pages or published scores.

Cut

442seconds per image

One turned a red dress green and took 442 seconds per image — cut.

Cut

One added a shoulder strap that doesn't exist on the real garment — cut, because a shop can't list that.

Fallback

3.7×slower

One matched the quality but ran 3.7× slower — kept as a fallback, not the default.

So each tool runs on whatever passed the comparison for its own job. Some cost more than the ones we cut. Speed and faithfulness to the product were worth more than the price difference.

Behind each tool — seven tools, seven sets of rules

Every tool runs on the same instruction-assembly system, but each has rules of its own that came from images that broke. This is what happens after you press generate.

Your photo sets the ceiling — and the limits we haven't solved

The photo you upload decides the outcome, and it is not only a pixel count — it is how much detail the AI can actually see. Measured on two Thai-print T-shirts: a sharp 1280px photo with large lettering held its shape 5 times out of 5, while a blurry 720px photo with small lettering held it 3 out of 5 — and garbled the Thai text on the shirt all 5 times.

Held its shape (runs out of 5)

  • 1280 px5/5
  • 720 px3/5

Wherever the source photo is unclear, the AI guesses. Upscaling still struggles with lace and fine woven patterns; the fix for now is to generate at high resolution from the start rather than enlarge afterwards.

White lace with a fine floral pattern on blackIllustration

Clothing (Product to Model · Virtual Try-On · same model) needs the neckline, sleeves and hem all visible. Crop small prints or lettering to one garment filling the frame — text too small to read in the source becomes new words. Front + back needs the back in the photo, or the back gets drawn from imagination.

Perfume, skincare and held products need the label facing the camera and the product filling the frame — the more of the photo it fills, the more accurate the label. Tiny text such as volume or ingredients can still drift; in testing the weight on a lipstick box changed digits, so check before every listing.

Lingerie and swimwear need to be laid open so straps, bows, lace and hooks show, on a background that contrasts with the piece — that photo is all the AI has. Ghost mannequin needs the neckline, shoulders and hem, and the same photo run twice can come back in a different shape, so we never judge quality from a single image.

The vendor's content filter occasionally rejects a result without saying why — 4 of 12 failed jobs so far. When that happens, credits are refunded automatically.

Ready for your first product photo?

Nothing to type, no card required — sign up and get 10 credits instantly

Start free nowOpen studio
Our storyAbout us

Photos by Divazus Fabric Store on Unsplash