From 287c18df090ce6dfa4367c61d74d1d163fbc198f Mon Sep 17 00:00:00 2001 From: IAMCCS Date: Wed, 7 Oct 2026 00:24:45 +0200 Subject: [PATCH] Remove obsolete R44 documentation --- docs/H3_FACE_MULTIFACE_R44_2026-10-06.md | 30 ------------------- docs/R44_4_VIDEO_EDITOR_RTX_FIX_2026-10-06.md | 19 ------------ .../R44_5_RTX_FACE_REFINE_MIXER_2026-10-06.md | 18 ----------- ...XER_VIDEO_EDITOR_AUDIO_SLAVE_2026-10-06.md | 21 ------------- ...R_METERS_FACE_PRESETS_RTX_UI_2026-10-06.md | 15 ---------- 5 files changed, 103 deletions(-) delete mode 100644 docs/H3_FACE_MULTIFACE_R44_2026-10-06.md delete mode 100644 docs/R44_4_VIDEO_EDITOR_RTX_FIX_2026-10-06.md delete mode 100644 docs/R44_5_RTX_FACE_REFINE_MIXER_2026-10-06.md delete mode 100644 docs/R44_6_MIXER_VIDEO_EDITOR_AUDIO_SLAVE_2026-10-06.md delete mode 100644 docs/R44_7_MIXER_METERS_FACE_PRESETS_RTX_UI_2026-10-06.md diff --git a/docs/H3_FACE_MULTIFACE_R44_2026-10-06.md b/docs/H3_FACE_MULTIFACE_R44_2026-10-06.md deleted file mode 100644 index b533e8a..0000000 --- a/docs/H3_FACE_MULTIFACE_R44_2026-10-06.md +++ /dev/null @@ -1,30 +0,0 @@ -# IAMCCS MiniMax H3 Face Delivery R44 — Multi Face / Crowd Safe - -## Purpose -The existing Face Delivery path refined one tracked subject. R44 keeps that exact path for `single` mode and adds an opt-in sequential `multi_face` mode for crowd/procession shots. - -## Runtime model -- Tracking for every candidate is performed against the untouched native visible frames. -- `select_index=0..N-1` chooses a subject once; the external H3 FaceRefine tracker then follows that subject by continuity instead of re-ranking faces every frame. -- H3 crop/refine is run sequentially per accepted subject. VRAM therefore remains bounded to approximately one face-refine pass at a time; runtime scales roughly with the number of accepted faces. -- Stitching is cumulative: subject 2 is composited onto the result containing subject 1, subject 3 onto subjects 1+2, etc. -- A conservative trajectory deduplicator rejects candidate tracks that converge on an already-refined person. -- `min_face_coverage` skips subjects detected in too little of the shot, avoiding expensive H3 passes for incidental/background detections. -- The original native audio tensor, temporal extent, LongVid context and motion state are never modified by the face passes. - -## Controls on IAMCCS_MiniMaxH3FaceDeliveryR38B -- `face_mode`: `from_shotboard` (default), `single`, `multi_face`. -- `max_faces`: 1–8, default 4. This is an accepted-subject target, not a VRAM batch size. -- `min_face_coverage`: default 0.10. In multi-face mode a candidate must have detector hits on at least this fraction of source frames. - -## Shotboard profile -`face_detailer_profile = multi_face` activates multi-face automatically when `face_mode = from_shotboard`. -The existing `balanced`, `small_faces`, `gentle`, `sam_face_mask`, and `wide_character_12gb` profiles retain single-face behavior. - -## Compatibility -- Compatible with historical H3FaceTrackCrop 6-output contract. -- Compatible with H3FaceTrackCrop 1.1.x 7-output contract (`frame_count` appended). -- `wide_character_12gb` is now accepted by backend profile validation instead of being silently normalized to `balanced`. - -## 12 GB recommendation -Start at 512×512, 4 steps, denoise 0.20, max_faces 3–4. Raising max_faces does not increase peak VRAM because passes are sequential, but increases wall-clock time nearly linearly. For a shot reporting 8 faces, use 4 first and raise toward 8 only if the background faces materially need refinement. diff --git a/docs/R44_4_VIDEO_EDITOR_RTX_FIX_2026-10-06.md b/docs/R44_4_VIDEO_EDITOR_RTX_FIX_2026-10-06.md deleted file mode 100644 index 5393616..0000000 --- a/docs/R44_4_VIDEO_EDITOR_RTX_FIX_2026-10-06.md +++ /dev/null @@ -1,19 +0,0 @@ -# IAMCCS R44.4 — Video Editor duration authority + RTX Exporter preflight - -## Video Editor - -- Audio hard cuts are now lane-local. A clip on A1 is not shortened because a clip starts later on A2/A3/A4/A5. -- Manifests previously shortened by the old global hard-cut rule are healed when the cut carries the IAMCCS system marker. -- Manual video/audio linking no longer pairs unrelated imports merely because they share a T01/T02 lane index. -- Generated H3 video/audio pairing uses take + clip index when available. -- The legacy `zz_iamccs_shotboard_video_editor_media_fix.js` no longer replaces the newer integrated server-media importer (2026-09-18+). This restores authoritative server probing for manual video duration and embedded audio. - -## IAMCCS Exporter PRO / NVIDIA RTX VFX - -- Added non-throwing RTX runtime preflight. -- Exporter UI reports whether the NVIDIA RTX VFX runtime is actually available instead of claiming RTX is active as soon as the checkbox is enabled. -- If RTX is requested but `nvvfx` is missing, the exporter logs an explicit warning and completes the export with original frames instead of aborting the entire delivery. -- Metadata/report distinguish `rtx_requested` from `rtx_effective`. -- `apply_rtx_vfx()` now imports/preflights the optional runtime before unloading Comfy models. - -Actual RTX VFX processing still requires the NVIDIA/Deno runtime. The fallback is intentionally not labelled as RTX output. diff --git a/docs/R44_5_RTX_FACE_REFINE_MIXER_2026-10-06.md b/docs/R44_5_RTX_FACE_REFINE_MIXER_2026-10-06.md deleted file mode 100644 index 97361fb..0000000 --- a/docs/R44_5_RTX_FACE_REFINE_MIXER_2026-10-06.md +++ /dev/null @@ -1,18 +0,0 @@ -# IAMCCS R44.5 — RTX preflight UI + Face Refine preset UX - -## RTX Exporter PRO -- IAMCCS never modifies the Deno repository. -- Runtime preflight now reports READY / NOT INSTALLED, runtime source, Python executable, nvvfx module path and VideoSuperRes availability. -- Added CHECK RTX INSTALLATION and OPEN OFFICIAL INSTALL GUIDE controls. -- Deno prepared runtime markers remain discovery-only; missing RTX falls back to native export. - -## Settings PRO / Face Refinement -- Face preset groups use a three-column aligned layout with responsive single-column fallback. -- Every preset has English production guidance and a hover tooltip. -- Face Size presets no longer change Face Detailer Profile, so explicit Single / Multi Face strategy is preserved. -- Face canvas/window validation now runs after Settings PRO overrides are resolved. - -## Video Editor / AudioBoard Mixer -Recommended editorial control route: -EditorTakeRoute cine_linx -> IAMCCS AudioBoard Mixer -> IAMCCS BusOut -> Video Editor cine_editor_linx. -The main EditorTakeRoute cine_linx connection to the Video Editor remains in place for take/shot truth. diff --git a/docs/R44_6_MIXER_VIDEO_EDITOR_AUDIO_SLAVE_2026-10-06.md b/docs/R44_6_MIXER_VIDEO_EDITOR_AUDIO_SLAVE_2026-10-06.md deleted file mode 100644 index d28e72e..0000000 --- a/docs/R44_6_MIXER_VIDEO_EDITOR_AUDIO_SLAVE_2026-10-06.md +++ /dev/null @@ -1,21 +0,0 @@ -# R44.6 — IAMCCS AudioBoard Mixer → Shotboard Video Editor Audio Slave - -The Mixer is the editorial audio master when `sync_policy = mixer_overrides_cine_linx`. - -Supported live slave state in Shotboard Video Editor: -- A1–A5 volume -- gain dB -- pan -- mute / solo -- normalize / bypass / reverb metadata -- effect chains -- master bus -- master gain - -Recommended graph: - -`EditorTakeRoute cine_linx -> AudioBoard Mixer -> BusOut -> Video Editor cine_editor_linx` - -Keep the original `EditorTakeRoute cine_linx -> Video Editor cine_linx` connection as generation/take truth. The editorial audio branch must not replace it. - -The Video Editor frontend exposes `_iamccsVideoEditorAudioSlave`; the Mixer discovers the downstream editor through graph links and mirrors its state live. Backend queue/render repeats the same authority so preview and final export agree. diff --git a/docs/R44_7_MIXER_METERS_FACE_PRESETS_RTX_UI_2026-10-06.md b/docs/R44_7_MIXER_METERS_FACE_PRESETS_RTX_UI_2026-10-06.md deleted file mode 100644 index 4172107..0000000 --- a/docs/R44_7_MIXER_METERS_FACE_PRESETS_RTX_UI_2026-10-06.md +++ /dev/null @@ -1,15 +0,0 @@ -# IAMCCS R44.7 - -## Mixer / Video Editor - -The Shotboard Video Editor remains an audio slave when driven by IAMCCS AudioBoard Mixer. R44.7 adds per-track A1–A5 and master meter data to the slave snapshot. Meter values are sampled from the editor audio peak data at the live playhead and include track fader/gain and master gain. - -Continuous volume/gain/master-gain edits no longer call a forced audio seek. HTML audio preview gain is ramped over successive animation frames to avoid zipper/click artefacts while dragging faders. Timeline scrubbing and explicit seeks keep the existing seek behaviour. - -## Face Refinement presets - -Quality, Face Size and Definition are independent contextual preset families. One preset can be active in each family at the same time. Their value domains are intentionally separated so changing one family does not silently invalidate another active family. A reset action clears all three highlights and restores the current IAMCCS Face Refinement technical defaults while preserving enable/mode/max-face policy. - -## RTX Exporter - -The RTX runtime panel keeps the same preflight/probe behaviour but uses compact action labels (`CHECK RTX`, `INSTALL GUIDE`) with full English tooltips so the controls remain legible at the normal node width.