←All work

Project · AI product photography

Studio-Grade AI Product Photography Pipeline

A few phone photos of a garment go in. A full set of studio-grade shots comes out — flat-lays, on-model, detail crops — in minutes, checked against the real product.

Gemini Vision + image gen Python Validation loop
The same garment, photographed on a phone and generated as a studio shot — after The same garment, photographed on a phone and generated as a studio shot — before Phone photo Generated

Drag to compare. Same garment — one a phone photo, one generated.

~90%
Lower cost per shot
Days → min
Per-product turnaround
10+
Shots per product
How it works

The pipeline, end to end.

Seven composable steps. Photos go in; a set of studio shots that's been checked against the real product comes out. The spec drives two generators that run in parallel, an optional branch handles posing, and a validation loop refuses to ship anything that doesn't match.

Inside each step

How every node works.

Each step is its own small tool with a clean input and output, so the pipeline is built from parts you can run, swap, or rerun on their own.

  1. 01 product_spec_extractor Vision

    Read the garment

    A vision model studies every phone photo at once and writes a full technical spec of the garment — fabric and drape, cut, pattern, hardware, trim, and a colour palette in proper fashion terms. That spec, not the photos, is what every later step generates against.

    A few phone photos ~600-word garment spec
    Technique
    Multi-image synthesis. Close-ups carry texture, stitching and hardware; full-length frames carry silhouette and drape — the model fuses them into one description.
    The thinking
    Generators drift when they work straight from pixels. Turning the product into precise words first gives every downstream shot a single source of truth to match, so a button stays a button and blush pink stays blush pink.
    Extracted spec
    Category 2-piece long set
    Material Plush, velvet-like fleece
    Pattern White & blush stripes, coral bows
    Details Notched lapel, lace trim, patch pockets
  2. 02 pose_extractor Vision optional

    Extract the pose

    An optional branch. Point it at any model photo and it writes a purely anatomical description of the pose — stance, spine, every joint angle — precise enough to rebuild from the text alone.

    A model photo An anatomical pose spec
    Technique
    Identity-stripped extraction. It describes only body mechanics in anatomical terms — abduction, flexion, rotation — and is explicitly barred from naming skin, face, clothing or background.
    The thinking
    Separating pose from identity turns poses into reusable text. Extract once, build a library, and apply the same pose to any model later — without re-shooting, and without carrying over who was in the original frame.
    Pose spec · identity-free
    Stance Seated, weight through pelvis
    Spine Upright, slight left rotation
    Right arm Elbow ~50°, hand to clavicle
    Legs Loose cross-legged
  3. 03 product_folded_generator Image gen

    Generate the flat-lay set

    From the spec plus the raw photos, it shoots the flat-lays a store actually needs: a clean retail fold, a relaxed styling layout, and a macro crop for fabric and stitch.

    Spec + raw photos 3 studio flat-lays
    Technique
    Multi-source, fabric-aware prompting. The text fixes how the material behaves — silk ripples, fleece bulks, cotton creases — while the photos lock the exact pattern scale, piping colour and button count.
    The thinking
    Feeding both the words and the originals is what stops hallucination. The three shots aren't decoration — they map to three moments for a buyer: the listing thumbnail, the styled scroll, and the zoom-in fabric check.
    Classic retail fold
    Retail fold
    Relaxed flat-lay
    Relaxed
    Texture detail crop
    Texture
  4. 04 model_in_pose_generator Image gen optional

    Put a model in a pose

    The other half of the pose branch. Hand it a model and a pose spec and it regenerates that same model performing the pose, keeping their face, build and hair intact.

    Model + pose spec A posed model
    Technique
    Identity-locked pose transfer. The prompt pins facial features, body type and hair to the reference while the body executes the described pose exactly.
    The thinking
    With poses stored as text, one model can be re-posed consistently across a whole catalogue — the same look, shot after shot, without booking the studio again.
    Held vs. changed
    Kept Face, body type, hair
    Kept Skin tone, identity
    Changed Pose, from the spec
  5. 05 product_on_model_generator Image gen

    Try it on a model

    The virtual try-on. Given the model, the product photos and the spec, it dresses the model in the garment at studio quality — keeping the model's own pose, or taking a pose from the branch above.

    Model + product + spec An on-model shot
    Technique
    Geometric lock. The body outline, limb thickness and the negative space between limbs are matched pixel-for-pixel; only the clothing is generated, draped with real tension and fold at the joints.
    The thinking
    Try-on usually distorts the body — garments swim or skin bulges. Anchoring the silhouette and changing only the fabric keeps the person real and the fit believable.
    The garment generated on a model
    On-model — body locked, only the clothing generated
  6. 06 product_validator Quality gate

    Score it against the real product

    Before anything ships, a second model grades each generated shot against the original photos and the spec — scoring three things out of ten and listing the exact problems it sees.

    Generated shot + originals Scores + issue list
    Technique
    LLM-as-judge. Three orthogonal axes — pattern/print, colour, construction — each with its own pass threshold (8.5 / 8.5 / 7.0) and specific, factual issues rather than vague notes.
    The thinking
    A shot can look professional and still be wrong. Splitting the verdict into three axes means a failure tells you which kind of fix it needs — and the structured output lets the next step act on it without a human reading it.
    Validation · scored vs. original
    Pattern / print 10/ pass 8.5
    Colour fidelity 10/ pass 8.5
    Construction 9/ pass 7.0

    All three above threshold → ships. Anything below → retouch.

  7. 07 image_retoucher Image gen

    Fix only what's off

    When a shot fails, the validator's issue list becomes the retouch instruction. The retoucher corrects exactly those problems, and the result goes straight back to be scored again — round and round until it passes.

    Shot + issue list A corrected shot
    Technique
    Reference-locked, surgical edits. Composition and perspective are held fixed and untouched areas stay pixel-identical; it changes what it's told to and nothing else.
    The thinking
    A broad “make it better” prompt invents new problems. Constraining the edit to the named issues is what makes the loop converge instead of wandering — generate, score, fix, re-score, ship.
    A corrected shot, re-scored against the original
    A corrected shot, re-scored against the original
The core idea

A quality gate that fixes its own mistakes.

The last two steps form a loop. Every generated image is scored against the real product; anything below threshold is sent back with a precise list of what's wrong, retouched, and scored again. Nothing reaches the catalogue until it passes on pattern, colour, and construction — so the pipeline can run unattended without quietly shipping a wrong print or an off colour.

  1. Generate
  2. Score
  3. Fix what failed
  4. Re-score
  5. Ship on pass
The output

One product, a full shot list.

From a single set of phone photos: flat-lays, on-model shots, and detail crops — as many variations as the catalog calls for, every one checked against the real product.

Studio flat-lay
Studio flat-lay
On-model
On-model
Detail crop
Detail crop
Why it matters

Photography stops being the bottleneck.

For a brand with hundreds of items, photography is usually what caps how fast products get online well. This turns it into a same-day step: shoot a product on a phone, get a full on-brand set back, and do it across the whole catalog without booking a studio for each drop. It runs on cents of compute per product, and every image is checked against the real thing before it ships.

A studio day shoots a few products. This shoots the catalog.

Have a catalog to shoot?

Tell me what you sell and how big the catalog is, and I'll come back with what a first batch looks like and what it would take.

Get in touch