Drama episodes

Standard of Iron: The Dead Company is a dialogue series filmed in the arena. An episode goes from script to finished film in five steps, all driven by files under scripts/drama/episodes/<episode>/:

Stages, which are the staged performances the cameras film, live in tools/arena/promos/dead_company/stages/*.stage.json. The arena loads them at startup (tools/arena/arena_drama_scenarios.cpp), so moving an actor or retiming a step needs no rebuild. SOI_DRAMA_STAGES points the arena at another directory.

Making an episode

EP=scripts/drama/episodes/ep01_killed_in_advance
OUT=artifacts/drama/ep01
QWEN=~/.local/share/soi-qwentts/bin/python   # torch cu128, qwen-tts, faster-whisper, speechbrain
CBX=~/.local/share/soi-cbx/bin/python        # torch cu128, chatterbox-tts installed --no-deps
$QWEN scripts/drama/perform.py $EP/lines.json --work $OUT/voicework --takes 3
$CBX  scripts/drama/convert.py $EP/lines.json --work $OUT/voicework
$QWEN scripts/drama/cast.py    $EP/lines.json --work $OUT/voicework --out $OUT/vo
python3 scripts/drama/plan.py $EP --vo $OUT/vo --out $OUT/plan --quality final   # or fast / preview
python3 scripts/drama/graphics.py $EP --out $OUT/graphics
scripts/promo-guarded-capture.py $OUT/plan/spec.json $OUT/clips --chunk 8
python3 scripts/drama/contact_sheet.py --plan $OUT/plan --clips $OUT/clips --out $OUT/sheets
~/.local/share/soi-tts/bin/python scripts/drama/assemble.py $EP --plan $OUT/plan \
    --clips $OUT/clips --vo $OUT/vo --graphics $OUT/graphics --out $OUT/ep01.mp4

1. Voices are performed, then recast.
   • perform.py asks Qwen3-TTS VoiceDesign (Apache-2.0) for several takes of every line. The instruction is the character's voices description plus the line's direct field ("sudden furious outrage, forgetting to whisper").
   • VoiceDesign acts well, but it invents a new voice on every call: speaker similarity between two calls was about 0.
   • convert.py runs each take through Chatterbox VC (MIT) into the character's designed reference voice. The delivery survives, and similarity rises to 0.5–0.8.
   • cast.py transcribes each take with faster-whisper, scores voice similarity with ECAPA and pitch spread, and keeps the best one.
   • Lines it flags (WER over 0.2) are usually spellings, numbers or Latin words; read the heard field before re-rendering.
   • Kokoro (voice.py) is kept for scratch reads only; it is flat.
2. Timeline. Every time in the edit is an expression resolved against the recorded lines: 12.5, line:s03_05, line:s03_05@end+0.3, mark:crack-1, end-2.
   • Stage steps carry an at expression too. plan.py resolves it into the step's t and rewrites the stage file.
   • Re-recording the voices therefore moves walks, turns and reveals along with the cuts. Expressive voices made episode 1 run 1:20 longer.
3. Gestures. plan.py writes <stage>.gestures.json beside each stage.
   • Each line plays a talking gesture on its speaker: talk by default for lines of 1.6 s or more, or the line's gesture in beats.json (point, shrug, nod, crossed, salute, chin, open).
   • Gestures never overlap, and actors that already have a routine (sitting, kneeling) are left alone. A scene with "gestures": false stays still.
   • The arena turns the cues into a showcase routine with wait: gaps.
4. Plan.
   • Each scene is one continuous performance. Every camera the edit cuts to is captured once, over the span it is used plus 0.35 s handles.
   • Static cameras creep: dialogue framings push in (push, m/s), and wides (fov 36° and up) drift sideways (drift).
   • Framings narrower than 40° get handheld sway (shake degrees; steady turns it off).
   • A track camera rides a group with focus-space offsets, and keys give an explicit camera move.
   • Quality presets are scope 2.35:1: final is 1920×816 at 48 fps, fast is 24 fps, preview is 1280×544.
5. Capture with the guarded runner. It needs a GPU display, so suspend blanking for the run (xset s off -dpms) and restore it afterwards.
6. Assemble.
   • Picture:
     • Cuts frame-accurately from the cut list. A 48 fps capture is averaged in pairs down to 24 fps, which gives a 180° shutter blur.
     • Applies the scene's grade (an ffmpeg filter chain), then film grain and a vignette, and letterboxes to 1920×1080.
   • Sound:
     • Everything is mixed in stereo. Voices get a little room (night, tent, open, camp).
     • Every bed, music and effect file is first brought to dialogue level, so an edit's db means "this far under the voices". cc0/... paths come from the CC0 library in artifacts/drama/cc0/lib (see its manifest.json for sources and licences).
     • Beds duck 5 dB and music about 7 dB under speech. The mix is set to -16 LUFS by one static gain and a limiter.
   • Subtitles sit in the lower letterbox bar.

Staging rules learned the hard way

• Spawn positions are snapped to cell centres that sit half a cell past the requested point. The loader asks half a cell short, so an authored [x, z] lands where it says. Walk destinations are not snapped.
• A walker turns to its direction of travel. To end facing someone, make the last leg of the walk point at them. turn (a centre wheel) rotates by a relative angle.
• Civilian garments (builder_tunic_roman, carthage_robes, roman_civilian_mantle, the Carthaginian civilian headwrap) render on a standing figure but vanish while it walks under the swordsman renderer. Dress the cast in gallic_tunic, iberian_tunic or armor_light_carthage.
• Unarmed leads use the caster body (k_caster_humanoid_asset). On the sword body they stand in a fists-up sword guard.
• Headgear, not hair. Primitive hair read badly; period headgear reads well. palla_veil, pileus, turban and skull_face are head attachments in render/equipment/helmets/hair_bun.cpp. In head-socket space the face is +z, the eyebrows sit near y -0.1 and the crown near y 1.35, about 1.4 units in radius. Anything skin-coloured reads as a bald head, so dye cloth strongly.
• Props shove actors. A unit spawned beside a prop is pushed off its tile, and a supply_cart at scale 0.9 is about 3 m long. Keep props well clear of the blocking and out of any eyeline you want to film.
• Review every cut, not the posters. scripts/drama/contact_sheet.py pulls one frame from the middle of each cut. Verify blocking from the NN_shot.camera.jsonl group positions and facings, not from a guess.
• Redo shots without recapturing everything. Capture a spec holding just the redone shots into clips/zz_redoN. The assembler takes the last match in path order, so a redo wins over the original take.
• Grade a scene with "grade" (an ffmpeg filter) on the scene in edit.json, for example to cut through river mist.
• There is no lip-sync or head turn. Gesture clips are showcase moves in animation/showcase_pose_manifest.cpp (clips 90–97). They are designed with scripts/preview-showcase-pose.py and checked after bake_creature_assets with humanoid_preview --clip showcase_gesture_talk. A full-body gesture overrides locomotion, so keep gestures off walking actors.

The cast

Costumes are renderer keys under troops/roman/dead_company/ and troops/carthage/dead_company/, with loadouts in assets/visuals/unit_equipment_loadouts.json and colours in render/humanoid/runtime/style_palette.cpp. troops/iron_sepulcher/skeleton_legionary is a skeleton in Roman kit. skeleton_legionary_bare drops the helmet and wears skull_face, a cartoon skull, because the skeleton body's own head is a featureless egg. In a stage, look takes a short name (felix, livia, tiro, severus, bostar, skeleton_legionary) or any renderer key.