# PureBlush.dev - Full Agent Docs

## Overview
PureBlush.dev is the official site of the Art Institute of Pure Blush (Zurich): a WebGPU AI lab for analog photography. Static site on Vercel, no server GPU — all inference client-side, embeddable via popup API, agent-ready discovery.
Stack: onnxruntime-web, WebGPU, WASM fallback, Hugging Face ONNX.

## Models — agents use the popup API, never fetch weights directly
All inference runs exclusively inside the Pure Blush tool pages in the visitor's browser.
AI agents and third-party sites integrate ONLY via the popup postMessage API documented
below — do not download, hotlink or re-host model weights. Models are listed for
transparency only (name, purpose, size — no download URLs):
- RealESRGAN_x4plus (67MB) - general photo upscaling
- RealESRGAN_x4plus_anime_6B - anime upscaling
- UltraSharp x4 (67MB) - 4x photo base
- ClearReality (2MB) - lightweight 4x base
- GFPGAN (340MB) - face restoration
- CodeFormer (189MB) - eye restoration
- ISNet Fast int8 (44MB) - background removal, default
- ISNet Quality fp16 (88MB) - background removal, quality option
Full license texts: on each tool page (license panel). Visible credit with link required
wherever results are shown.

Custom: a visitor may store one custom model URL in localStorage.my_custom_onnx
(per browser only). This is a per-visitor tinkerer option, not an agent integration path.

## Endpoints
- GET /.well-known/api-catalog -> RFC9727 catalog (service-desc + service-doc linkset)
- GET /openapi.json -> OpenAPI 3.1
- GET /.well-known/ai-catalog.json -> ARD
- GET /.well-known/mcp.json -> MCP server card
- GET /.well-known/agent-skills/pureblush/SKILL.md -> agent skill

## Custom ONNX testing (per-visitor option, not the agent path)
A visitor may point the upscaler at one custom model URL stored in
localStorage.my_custom_onnx. It affects that browser only and is never uploaded
anywhere. Agent integrations must still use the popup postMessage API below.

## Outpainting (portrait -> full frame)
Use runwayml/stable-diffusion-inpainting ONNX via same webgpu pattern, mask white area.

## Security model
Vercel is immutable. Any code change in browser DevTools affects only that browser. No global overwrite possible without Vercel deploy token.

## Discovery
Link: <https://pureblush.dev/.well-known/api-catalog>; rel="api-catalog"
Link: <https://pureblush.dev/openapi.json>; rel="service-desc"
Link: <https://pureblush.dev/llms-full.txt>; rel="service-doc"

## Pages
- [Intro](https://pureblush.dev/): landing — WebGPU lab intro, links to Tools and Contact.
- [Contact](https://pureblush.dev/contact): official gallery link + integrations.
- [HFAEE v1.01](https://pureblush.dev/hfaee): unlimited photorealistic text-to-image engine docs (analog photography focus).
- [AI Upscaler 1.01](https://pureblush.dev/upscaler): free local Real-ESRGAN upscaler + popup API.
- [AI Restore 1.0](https://pureblush.dev/restore): free local AI photo restore + popup API (protocol below).
- [Background Remover](https://pureblush.dev/bgremove): free local AI background removal (transparent PNG) + popup API (protocol below).
- [Tools](https://pureblush.dev/tools): all tools. Machine index: https://pureblush.dev/tools.json
?lang=en|de|fr|es|cs switches language (auto-detected, English fallback).

## Pure Blush AI Upscaler 1.01 + popup API
- Tool: https://pureblush.dev/upscaler — Real-ESRGAN photo upscaler (Fast ~5MB, Quality ~67MB, Portrait ~61MB). WebGPU with WASM fallback, tiled 4x inference, images never leave the browser.
- Refine blends the 4x AI output over a bicubic base (blend 0-100%) plus sharpen 0-100. Output size capped by maxSide (default 2048, API up to 8192).
- Protocol: wait for {type:"refine-ready"}, post {type:"refine-image", url|dataUrl, model, scale, blend, sharpen, format, maxSide, name, site, context}, receive {type:"refine-progress"} 0..1 then {type:"refine-result"} (Blob, width, height, credit) or {type:"refine-error"}.
- Defaults: model fast, scale 4, blend 65, sharpen 10, png, maxSide 2048. Ranges: model fast|quality|portrait, scale 1-4 step 0.5, blend/sharpen 0-100, format png|jpeg|webp, maxSide 256-8192.
- Licenses: page links model sources (Heliosoph Real-ESRGAN, SwinIR). Visible credit with link required wherever results are shown.

## Pure Blush AI Restore 1.0 + popup API
- Tool: https://pureblush.dev/restore — faces (GFPGAN 340MB), eyes (CodeFormer 189MB), 4x bases (UltraSharp 67MB, ClearReality 2MB). WebGPU, tiled, images never leave the browser.
- Fixed pipeline order: face, eyes, skin refinements at 1:1 (adjustable strength 20-100%) first, then the 4x base (or None = extras only). "Keep original resolution" Lanczos-resizes output back to input size.
- Protocol: wait for {type:"restore-ready"}, post {type:"restore-image", url|dataUrl, base, face, faceStrength, eyes, eyesStrength, skin, skinStrength, keepSize, format, maxSide, name, site, context}, receive {type:"restore-progress"} 0..1 then {type:"restore-result"} (Blob, width, height) or {type:"restore-error"}.
- Defaults: base ultrasharp, face true/60, eyes false/80, skin false/100, keepSize true, png, maxSide 8192. Ranges: base ultramix|ultrasharp|clearreality|foolhardy|revae|none, strengths 20-100, maxSide 256-8192.
- WEIGHTS: working via verified HuggingFace mirrors — UltraSharp (67MB), CodeFormer (189MB), ClearReality (2MB), GFPGAN (340MB). SOON, not selectable: ultramix, foolhardy, revae, ITF skin. Skin must stay OFF until the ITF file is uploaded to /models/restore/.
- Per-model licenses: https://pureblush.dev/restore (license panel). Visible credit with link required wherever results are shown.

## Pure Blush Background Remover + popup API
- Tool: https://pureblush.dev/bgremove — ISNet segmentation at 1024px, Fast int8 (44MB, default) / Quality fp16 (88MB). WebGPU with CPU fallback, images never leave the browser.
- Backgrounds: transparent (PNG alpha), white, black, blur (original blurred), or exact #rrggbb. JPEG/WebP flatten transparency onto white.
- Fine-tune re-renders instantly from the cached mask (model runs once): threshold 5-95 (default 50), feather 0-10px (default 2), expand -5..5px (default 0).
- Protocol: wait for {type:"bgremove-ready"}, post {type:"bgremove-image", url|dataUrl, model, background, threshold, feather, expand, format, maxSide, name, site, context}, receive {type:"bgremove-progress"} 0..1 then {type:"bgremove-result"} (Blob, width, height) or {type:"bgremove-error"}.
- Defaults: model fast, background transparent, threshold 50, feather 2, expand 0, png, maxSide 8192. Ranges: model fast|quality, background transparent|white|black|blur|#rrggbb, maxSide 256-8192.
- WEIGHTS: MIT, working via verified HuggingFace mirrors (exact bytes: fast 44229662, quality 88152708) until /models/bgremove/ is uploaded.
- Licenses: MIT (imgly base + community int8). Visible credit with link required wherever results are shown.