# wave Enhance

AI super-resolution for video — rendered in-graph by FFmpeg 9's ONNX DNN backend.

- This host (enhance.wave.online) is a WAVE Protocol Plane spoke: a thin edge front-door.
- The API, auth, scopes, metering, and pay-per-use all live at the gateway: https://api.wave.online
- Errors are JSON: {"error":{"code","message"}}. Authenticate with a wave key or pay via x402.

## Surfaces
- [Landing page](https://enhance.wave.online/)
- [Status](https://enhance.wave.online/status) — health + dependencies (append ?format=json)
- [Transparency](https://enhance.wave.online/transparency) — data + logging policy
- [Health](https://enhance.wave.online/health) — liveness probe (JSON)
- [llms.txt](https://enhance.wave.online/llms.txt) — this document
- [Structured manifest](https://enhance.wave.online/index.json) — index.json
- [Atom feed](https://enhance.wave.online/feed.xml) — product updates

## For agents
- [MCP manifest](https://enhance.wave.online/.well-known/mcp.json) — tool discovery (server hosted at the gateway)
- [Agent card](https://enhance.wave.online/.well-known/agent-card.json) — A2A
- [x402 discovery](https://enhance.wave.online/.well-known/x402) — pay-per-use
- [did:web identity](https://enhance.wave.online/.well-known/did.json)


## Enhance v1 constraints
- Model: ESPCN super-resolution (Apache-2.0). Output is ALWAYS a fixed 672x672 (3x of the model's
  fixed 224x224 working canvas) — verified empirically that FFmpeg 9's ONNX backend enforces this
  model's compiled-in shape; every input frame is letterboxed to 224x224 first, not tiled/proportional.
- CPU execution provider only in v1; GPU EP (`device=cuda|dml|vitisai`) is a named fast-follow.
- Input capped at 1280x720 (a product-quality/cost-sanity guardrail, not a latency workaround —
  inference cost is constant regardless of input size once letterboxed to 224x224).


[a WAVE product](https://wave.online)