mirror of
https://github.com/ruvnet/RuView.git
synced 2026-10-22 06:32:35 -05:00
Fifth style preset for the ADR-351 room-visualization pipeline: matches RuView own marketing hero image via fal-ai/fast-sdxl/image-to-image (base64 data-URI image_url, no separate storage upload needed - confirmed via a real minimal test call before implementation). Unlike the other four presets, this one deliberately depicts a literal human pose-skeleton figure, matching the reference image aesthetic - a materially weaker privacy posture than the abstract preset, documented explicitly in ADR-351 rather than blurred. Real 8-frame capture against the live household feed produced a real, ffprobe-verified H.264 MP4. All 8 frames are byte-identical (same MD5) - checked honestly: raw sensing features varied between live polls during capture, but the coarse Activity bucket build_prompt consumes stayed at OccupiedMoving for the whole window, so deterministic fast-sdxl output given fixed seed/prompt/reference/strength is expected, not a bug. Real spend: 9 fal.ai calls (1 connectivity test + 8 capture frames), well under the 20 dollar budget for this work.