🥺 Please, just 2 minutes — we beg you to read this first
😭 We poured everything into this. Please don't skip — 2 minutes is all it takes to understand what it actually does, and it changes everything.
1 Watch the demo
2 This is not an image-to-image transform
Every other AI cosplay tool paints a costume onto your photo. You get a hero you can look at and never wear. We output the plan, not the picture — the render is just proof the plan lands.
3 What YouCam is — and where it stops
Knowing what kind of API this is decided what we could build on it. It is an async measurement-and-try-on API, not a generative one — so the intelligence has to live in what you do with the measurements.
Task-based, not one call
Every feature is upload → presigned PUT → start task → poll for the result. Nothing is synchronous image-in / image-out.
Measurement, not generation
Skin AI returns numbers — redness, radiance, oiliness, a tone hex. It reads the person in front of it; it does not invent one.
Garment-scoped rendering
Apparel VTO swaps clothing on a real body and preserves pose and lighting. It is a try-on engine, not a character generator.
Gaps we hit, and what we did about each
The two APIs disagree about framing
Skin AI wants a face filling the frame. Apparel VTO wants head and shoulders. One photo cannot satisfy both.
↳ 4-step adaptive crop ladder — on our sample photo, attempts 1-4 are rejected and the 5th succeeds
Undocumented error codes
error_src_face_too_small and error_large_face_angle come back live but appear in no doc page.
↳ Both mapped to plain-English guidance instead of a generic failure
No-face photos fail slowly
A DLQ 'max retries exhausted' after ~70s rather than error_no_face, which is long enough to blow a serverless budget.
↳ Poll timeouts capped by hand at 20s skin / 45s cloth
Live payloads differ from the docs
skin-analysis takes a flat dst_actions body, not file_sets/actions. Polling is a path segment — ?task_id= returns 405. v2.0 cloth upload requires file_size; v1.0 does not.
↳ Every shape re-derived against the live API and written up in the README
HD and SD action names cannot mix
One request must be entirely HD-prefixed or entirely SD, or it is rejected.
↳ Locked to the SD set
Garments only — no props or accessories
VTO renders clothing. Masks, capes, emblems and props are exactly what makes a cosplay read as a character.
↳ Those become written guide steps, so the plan covers what the render cannot
4 How it is wired
4 YouCam task APIs · full run 90–170s
Your photo
One shot, in clothes you already own
EXIF-rotate · downscale to 2560px · flatten to white · re-encode (sharp)
↓Skin AnalysisYouCam Skin AI
redness · radiance · acne · oiliness
Skin Tone AnalysisYouCam Skin AI
skin · eye · lip · brow colour → 0–10 score
Run concurrently — serialising them pushed a measured run to 2m48s. Either can fail without failing the request.
↓Hero match
Tone score weights which hero you get in auto mode
The skin reading changes who you become, not just what the caption says
↓Guide written
Measured skin facts ground the makeup section · strict JSON out
Gemini writes the prose around what YouCam measured
↓Text-to-ImageYouCam
Garment generated per run from the guide's own prompt
So the costume matches your character instead of being one of a few fixed files
↓Apparel VTOYouCam
Costume rendered onto your real body — pose and lighting preserved
What you get
Before/after · costed steps · makeup · pose guide · PDF
🤔 So why is Gemini in there at all?
Because YouCam has no endpoint that reads an outfit and writes a plan — that is not what the API is for. Gemini fills exactly that gap: it looks at your clothes, takes the Skin AI numbers as given facts, and returns strict JSON. It is deliberately not the core. It writes prose around what YouCam measured. We swapped it from GPT-4 Vision mid-build and the output was materially the same — that is how replaceable this slot is. What is not replaceable is the YouCam layer: it decides who you are matched to, what garment exists, and what you end up looking like.
💄 The makeup try-on we built and then deleted
An early version rendered makeup onto the face. We cut it — makeup try-on is a separate Perfect Corp product, outside the Skin AI + Apparel VTO set this hackathon is built around, and leaning on it would have meant scoring points with an API we were not asked to use. Makeup survives as written guidance instead, grounded in the real Skin AI scores rather than rendered.
5 What's missing — and why it would be a game changer
Everything below is something we ran into building this, not a wishlist from the docs. If the YouCam API shipped these, this stops being a guide with a picture attached and becomes the actual product — you would see the finished cosplay, not an approximation of it.
🩹Add to a garment instead of replacing it
today To glue one felt spider onto a shirt, we have to generate a whole replacement outfit and swap everything the person is wearing.
if it existed A DIY guide is a list of small additions. One call that says 'keep this garment, add this shape at the chest' would make the render match the written steps exactly, every time.
🎭Props, masks and accessories
today VTO renders clothing only, so the mask, cape and emblem stay as written instructions the user has to imagine.
if it existed A cosplay is recognisable because of the mask and the emblem. Rendering those is the whole distance between 'a red shirt' and the character.
🧵Texture fidelity in the render
today Hand-cut felt edges and fabric-paint brush strokes are clearly visible in the reference image, then smoothed away once it lands on a body.
if it existed The promise is 'you made this yourself'. Keeping the material texture is what makes a homemade result actually look homemade.
📸One photo, both products
today Skin AI wants a face filling the frame, Apparel VTO wants head and shoulders. We bridge it with a 4-step crop ladder that can burn ~90s in retries.
if it existed A single 'analyse this person' mode that accepts one wider photo and crops internally would delete the hardest part of combining these two products.
💄Makeup rendering inside this API set
today We built a makeup try-on and deleted it — it belongs to a different Perfect Corp product, outside the set this hackathon is built on.
if it existed We already measure the skin. Rendering the makeup we recommend, on the same face, would close the loop from measurement to result.
CosplayAI
Transform Into Your Hero
$15. 40 minutes. Zero sewing skills required.
AI-powered cosplay guide that works with clothes you already own
🎭 How should we pick?
📋 Photo Requirements:
- ✓ Face + shoulders visible
- ✓ Standing position
- ✓ Good lighting
- ✓ Any background (we'll remove it!)
