Imagehide demo

Embed 896 invisible bits into any image, then recover them after a JPEG round-trip, a resize, or a social-platform pipeline. Everything runs on your device.

Highlight any text → 💬 comment. Publishes to GitHub Issues; optional 🤖 Ask Claude to fix tag for action items.

loading model…

Encode

Embed a payload into a photo.

Cover image

Drop a photo, or browse

or use the sample

cover

Payload

Auto — pHash + Ed25519 sig + demo pk Custom — type a message (up to 96 chars)

0 / 96 bytes

Loading model…

—

PSNR dB

—

SSIM

—

encode time · 100–500× in prod

container (watermarked)

residual ×10

Show payload bits

Decode

Recover the payload — optionally through a channel attack.

Image to decode

Last container — none yet, encode one first Upload — any image with an embedded payload

Drop, or browse

image

Channel attack (applied before decoding)

Waiting for an image…

—

bit accuracy

—

signature

—

rs ecc · bytes

—

phash drift · bits

—

decode time · 100–500× in prod

Show recovered bits

About this demo

What it is. Live demo of the MWIP watermark — an invertible neural network embeds an 896-bit payload (H = 128-bit perceptual hash, sig = 512-bit Ed25519, pk = 256-bit public key) into an image, and a sibling network recovers it after channel distortion.

Demo caveat. The keypair ships in this page; anyone can forge a valid signature. Real deployment uses device-bound secret keys. The pHash is a placeholder.

Where this runs. ONNX Runtime Web in a Web Worker, on your device. Your image is never uploaded. Out of scope for the current model: crop, rotate, Gaussian noise, Gaussian blur.

Image size cap. The decoder allocates ~12 MB of float32 per megapixel plus ~60–120 MB of forward-pass intermediates. The demo caps input at 1 MP on desktop (WebGPU) and 0.5 MP on mobile (WASM on iPhone, where Safari's per-tab memory budget is tight). The model itself has no such limit — on a phone NPU or server it runs the full image.

Production speed per pass (encode ≈ decode — same INN backbone):

Photo size	Pixels	Phone NPU	Server GPU fp16	Server CPU AVX-512
Instagram post	1080×1080 (1.2 MP)	~80 ms	~7 ms	~30 ms
1080p screen capture	1920×1080 (2.1 MP)	~140 ms	~12 ms	~55 ms
iPhone screenshot	1179×2556 (3.0 MP)	~200 ms	~17 ms	~80 ms
16″ MacBook screenshot	3456×2234 (7.7 MP)	~510 ms	~45 ms	~200 ms
Phone camera (default)	4032×3024 (12 MP)	~1.0 s	~70 ms	~300 ms
Phone HQ / iPhone 24 MP	5712×4284 (24 MP)	~2.0 s	~140 ms	~600 ms
Full-frame DSLR	7008×4672 (33 MP)	~2.7 s	~190 ms	~830 ms

Browser-WASM mobile is ~100–500× slower than the phone-NPU target; server GPU is ~10× faster again.