tailor·megithub.com/…/techpack-vectoriser ↗
Index — I

An AI tailoring tool,
built openly.

Photo of a garment in, tech-pack vector out — and eventually, a way for anyone to tailor clothes they like online without guesswork.

Try it  →or read the roadmap

Plates I–IV — from the new engine

04 of 04 · front + back
PL. I t-shirt · front + back01 of 04
tshirt front

front

tshirt back

back

From a phone photo of a plain crew tee. Set-in sleeves with cuff topstitch, neckline binding, hem topstitch — all traced as vector paths.
PL. II hoodie · front + back02 of 04
hoodie front

front

hoodie back

back

Pullover hood with drawstring, kangaroo pocket, ribbed cuffs and waistband. Same engine, taken from a single photo.
PL. III jacket · front + back03 of 04
jacket front

front

jacket back

back

Bomber with full zip, welt pockets, sleeve pocket, ribbed collar / cuffs / hem. Recognised and drawn from one image.
PL. IV dress · front + back04 of 04
dress front

front

dress back

back

V-neck flutter-sleeve A-line. Waist seam, belted back, hem topstitch. The engine handles tailored garments too.

Roadmap

II

What is, what is being made, and what is to come.

  1. 01

    Photo → tech-pack flat → SVG

    End-to-end engine. Photo of a garment in, professional black-line flat sketch out, traced into editable vector paths. Runs in ~7 seconds per view via Modal + Replicate. Verified on t-shirt, hoodie, polo, jacket, dress.

    Shipped
  2. 02

    Public /try upload

    Drop any garment photo in the browser and get the flat back. Wired directly to the production engine — same pipeline that produces the plates above.

    Shipped
  3. 03

    Front + back from a single photo

    Two views generated from the same upload: the front view of the garment as photographed and an inferred back view in the same construction language. Together they form a starter tech pack.

    Shipped
  4. 04

    Side view (needs different architecture)

    Off-the-shelf image models persistently draw a front when asked for a side profile. The honest path is 3D garment reconstruction → orthographic projection, not more prompt engineering on a 2D model. Parked until that's viable.

    Later
  5. 05

    Measurement callouts + factory pack

    Thin leader lines labelling sleeve, neckline and hem on the output SVG, plus a millimetre rule along one edge. Optional metadata layer for production handoff.

    Building
  6. 06

    Fine-tuned model on a hand-drawn flats dataset

    The real moat: build a paired (photo, hand-drawn-tech-pack-SVG) dataset and fine-tune a FLUX LoRA on it. Beats the off-the-shelf vision model on hard cases — tailored, technical, couture.

    Later
  7. 07

    Tailor-to-photo measurements

    Inverse problem: given a photo and a target body, produce the alteration deltas. The eventual goal — easier online tailoring for anyone.

    Later
III

Recent commits

Fetched from the public repository. Build in public means the broken commits show up too.

View on GitHub ↗

main · updated 3 weeks ago

  • 9d648b3wip: snapshot (17 file change(s) idle 5675 m at 2026-06-17 15:48)3w
  • fe59a43scripts: Rodin multi-view side-render + tight-crop retrace4w
  • 4eb38a1scripts: Hunyuan3D side-render test wrapper + SVG normalisation4w
  • ed7c490scripts: silhouette-to-filled-SVG converter for inline-tint customisers4w
  • 2f15d50v3.22: positively-described side-profile prompt (no "rotate this" language)4w
  • bc08219v3.21: replace pixel-zone whitener with component-based crease removal1mo
  • 73d456ev3.20: scan + targeted remediation for creases on heavy garments1mo
  • a5e3ad1v3.19: third FLUX polish pass on heavy garments (hoodie, sweatshirt, jacket, sweater)1mo
IV

Bring a photo.

A clean front view of a t-shirt, hoodie, jacket, or polo. Roughly ten seconds for the front; about two minutes if you'd like a side.

Try it  →