S — SOUTH

Lumen

Lumen

Version 1.00Source on GitHub

DEPLOYMENT

Quality deployment solutions — dotfiles, configuration, and release tooling that carry a project from first checkout to production.

HOSTeditor · agent appthe user works hereapproves tool callsMCP CLIENTone per serverspawns the serverframes the messagesMCP SERVER — THIS PAGEtools · resources · promptsdoes the work, returns resultsnever reaches past its contractstdioTHE SESSION, IN ORDER1 · initialize →protocol version + capabilities exchanged2 · notifications/initialized →the client is ready3 · tools/list → ← tool schemasnames, descriptions, input schemas4 · tools/call → ← resultthe work itself, as structured contentEVERY MESSAGE IS A JSON-RPC 2.0 ENVELOPE
The Model Context Protocol session every MCP server on this page speaks: initialize, tools/list, tools/call — JSON-RPC 2.0 envelopes over stdio.
Before and after panels of Lumen style grades applied to a single portrait crop
Lumen's shipped grade presets applied to one portrait crop by the engine itself — the before/after board from the 2026-10-18 lightshow video study, produced with the engine's own style operations.
The demo_toon grade checked on a pixel-art sprite
The demo_toon preset — landed as a shipped style on 2026-10-18 — checked on a pixel sprite from the same study: the posterize step that speckles painted portraits stays clean on pixel art.

BENCHMARKS

Lumen — style grade throughput (measured 2026-10-18)
InputPresetMedian per gradeThroughput
448×448 portrait cropdemo_toon_portrait7.77 ms25.8 Mpix/s
448×448 portrait cropukiyo_e8.23 ms24.4 Mpix/s
1024×1024 (Lanczos3 upscale)demo_toon_portrait41.0 ms25.6 Mpix/s
1024×1024 (Lanczos3 upscale)ukiyo_e42.6 ms24.6 Mpix/s
Measured on primo (13th Gen Intel Core i7-13620H) with Lumen's release build, timing lumen::style::grade_image — the pixel path behind the style_apply MCP tool — at strength 1.0: 20 timed runs per row after 3 warm-ups, the median reported; PNG decode/encode and the per-run input clone sit outside the timed region. The subject is the 448×448 Séraphine portrait crop from the 2026-10-18 Lumen study, and that crop upscaled to 1024×1024. Throughput counts input megapixels per second. Raw record: run_lumen_bench.py and lumen-results.json in the site-bench-20261018 work record.

REFERENCES

  1. Model Context Protocol — specificationLumen's style and VRM operations are served as MCP tools (stdio).
  2. A2A Protocol — Agent2Agent specificationLumen also serves the same operations over the A2A agent protocol.
  3. VRM specification — VRM consortiumThe avatar format Lumen inspects and grades: a glTF 2.0 extension for humanoid avatars.
  4. glTF 2.0 — Khronos specificationThe base 3D format under every VRM file: the GLB container Lumen parses.