Measured with RawNIND on a 20.8 MP raw: an Apple M3 does a tile in 116 ms, an
RTX 3060 Ti over Tailscale 345 ms there and back; sharing blindly made the
export take 34 s instead of 6.4 s while the photo waited for the server's last
tiles. The greeting now times a check tile there and back, and tiles are
shared only when that beats this computer's own time per tile (the card's, or
the CPU's with its threads counted as `card_is_faster` does). Otherwise
`denoise.status` → `device.remote` says `slower` with both times. With the
server configured but slower the export takes 7.2 s (greeting + check tile).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- `lightcraft_denoise::remote`: protocol v1 over TCP (JSON lines + byte-shuffled,
deflated half-float tiles), a `Remote` client (connection pool, four slots,
greeting checks the model's SHA-256 and tile size, gives up after three failed
tiles) and `serve` (shared token compared in constant time, size caps including
payloads that inflate past their tile, connection limit, timeouts).
- Engine: with LIGHTCRAFT_DENOISE_URL / LIGHTCRAFT_DENOISE_TOKEN set, the loaded
model greets the server on first use and adds its slots to this computer's
tiles at once; a tile goes to a free server slot, else runs here, and a tile
the server fails runs here too. `denoise.status` → `device.remote` says where.
- `lightcraft-cli denoise-serve` serves the installed, selected model on this
computer's card (checked against the CPU, as in the app) or CPU.
- docs/denoise.md: setup, safety (unencrypted: VPN/LAN only) and the protocol.
Tested: protocol round trips, refusals (token, model, tile size, hostile
payloads), fallback and give-up; the engine model with a live, a mismatched
and a dead server; the CLI binary serving a synthetic model end to end.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Recover the unpublished denoise work against current upstream. Preserve later opt-in model setup, shared Rust HTTPS downloads, pending-photo resume and an Amount-independent switch. Keep original RAWs unchanged and share the current color finish between ordinary and denoised sources.
Bound public tiling inputs, refuse corrupt-cache exports, and retain synthetic/real-model coverage. Include two Windows path spelling fixes in existing tests. No model weights, face code, media or Tract runtime are included.
Validation: cargo xtask ci (all seven gates); installed RawNIND CPU/GPU agreement 1.38e-6; native CPU preview/export on a public 24 MP Sony RAW, unchanged original SHA-256.
Co-authored-by: Codex <[email protected]>
Co-authored-by: Claude Sonnet 5.5 <[email protected]>
Generated with Codex.
