patterns.yaml

# The pattern library: tried-and-tested recipes the pattern chooser
# filters. Each pattern is one preset that routes a piece of work to a
# known-good workflow. Edit here, run `python -m engine site`, and the
# chooser page regenerates.
#
# Fields:
#   lane      recon | release | hero   (may list several)
#   stage     analyse | cast | panels | motion | scene | video | assemble
#   tier      draft | standard | premium
#   status    built | designed | planned  (honest state, shown as a badge)
#   tags      capability chips, also filterable
#   models    which providers.yaml families this leans on (provisional
#             until the next dated control-stack research pass)
#   cost      rough AUD (AUD/USD 0.70, 2026-07-20); "-" when negligible
#   inputs / outputs / risk   shown when a card is expanded

patterns:
  - id: lyric-analysis
    name: Lyric analysis pass
    lane: [recon]
    stage: analyse
    tier: draft
    status: built
    tags: [text-only, gate]
    summary: Turn lyrics into themes, mood, story seed and 8-12 panel beats before any spend.
    models: [claude CLI or local Qwen]
    cost: "-"
    inputs: A vault note at status ingested.
    outputs: Analysis merged into the note; status becomes analysed.
    risk: Cheap to redo. Read the story seed and beats and fix by hand.

  - id: recon-draft-clip
    name: Recon draft clip
    lane: [recon]
    stage: video
    tier: draft
    status: designed
    tags: [batch, harvest-lora-data]
    summary: Fast image-to-video test on the cheapest tier; keep the good frames as LoRA training data.
    models: [HunyuanVideo 1.5]
    cost: "≈A$0.03-0.10 per 6s clip"
    inputs: A keyframe or panel plus a short motion prompt.
    outputs: A rough clip; salvaged stills feed the cast foundry.
    risk: Accept jank. Batch overnight on interruptible GPUs; cap spend per song.

  - id: model-bakeoff
    name: Model bake-off
    lane: [recon]
    stage: video
    tier: draft
    status: designed
    tags: [batch, evaluation]
    summary: Run the same shot across several models to see which wins on your own material.
    models: [HunyuanVideo 1.5, Wan 2.2, LTX-2.3]
    cost: "≈A$0.10-0.30 per shot per model"
    inputs: One fixed shot spec.
    outputs: A side-by-side sheet; the winner gets promoted for that use.
    risk: Only promote a new model to Hero after it beats the incumbent here.

  - id: character-lora
    name: Character LoRA foundry
    lane: [recon, hero]
    stage: cast
    tier: premium
    status: designed
    tags: [char-lora, one-off, asset]
    summary: Train a reusable LoRA for a hero recurring character; validate before it's allowed downstream.
    models: [FLUX.1-dev or Qwen-Image base]
    cost: "≈A$2-7 per character (30-60 min A100)"
    inputs: A curated reference set, ideally seeded from your own art and photos.
    outputs: A versioned .safetensors in the asset registry, with trigger word and sample grid.
    risk: A bad LoRA poisons every downstream job. Test-sheet it before use; version it.

  - id: singer-band-lora
    name: Singer / band likeness LoRA
    lane: [recon, hero]
    stage: cast
    tier: premium
    status: designed
    tags: [char-lora, likeness, asset]
    summary: Same foundry, aimed at real people (singers, band, supports) for on-model performance shots.
    models: [FLUX.1-dev or Qwen-Image base]
    cost: "≈A$2-7 per person"
    inputs: Consented reference photos of the person.
    outputs: A versioned likeness LoRA in the registry.
    risk: Keep consent and any private likenesses local; mark the asset private.

  - id: reference-character
    name: Reference-image character
    lane: [recon, release]
    stage: cast
    tier: standard
    status: designed
    tags: [char-ref, no-training]
    summary: Hold a character across shots with multi-reference, no training. Good for occasional cast.
    models: [FLUX.2 multi-ref, Qwen-Image-Edit]
    cost: "generation cost only"
    inputs: 1-10 reference images of the character.
    outputs: On-model frames without a trained LoRA.
    risk: Weaker lock than a LoRA under heavy motion; fine for lighter appearances.

  - id: comic-panels
    name: Comic pre-viz panels
    lane: [recon, release, hero]
    stage: panels
    tier: draft
    status: designed
    tags: [stills, cheap-gate]
    summary: Render 8-12 stills from the panel beats. The second cheap gate before any video spend.
    models: [Qwen-Image, FLUX.2 Klein]
    cost: "≈A$0.01-0.05 per panel"
    inputs: A briefed note with panel beats and its cast.
    outputs: A panel set; status becomes panels. Approve the strongest.
    risk: Stills are cheap to re-roll; do the visual arguing here, not in video.

  - id: keyframe-select
    name: Keyframe selection
    lane: [release, hero]
    stage: panels
    tier: draft
    status: planned
    tags: [selection]
    summary: Promote the approved panels to start and end frames for video shots.
    models: []
    cost: "-"
    inputs: Approved panels.
    outputs: Keyframe pairs that bookend each motion.
    risk: Pick pairs that imply believable motion between them.

  - id: pose-capture
    name: Pose capture + retarget
    lane: [release, hero]
    stage: motion
    tier: standard
    status: designed
    tags: [pose-capture, control]
    summary: Capture a move from any footage and retarget it onto your character.
    models: [DWPose, AnimateAnyone/UniAnimate-class]
    cost: "capture is local and cheap; drive is video-gen cost"
    inputs: A reference clip (phone footage, a performance) plus a target character.
    outputs: A pose-driven clip of your character performing the move.
    risk: Capture and pose maps run local; only the driven render costs GPU.

  - id: beat-sync-motion
    name: Beat-synced motion
    lane: [release, hero]
    stage: motion
    tier: standard
    status: designed
    tags: [beat-sync, control]
    summary: Snap a captured pose track to the song's beat grid so the movement lands on the music.
    models: [librosa, Demucs]
    cost: "-"
    inputs: A pose track and the song's audio.
    outputs: A beat-aligned motion track ready to drive generation.
    risk: Runs local on audio analysis; no GPU cost until render.

  - id: choreo-ops
    name: Duplicate / mirror choreography
    lane: [release, hero]
    stage: motion
    tier: standard
    status: planned
    tags: [control]
    summary: Apply one captured motion to several cast members, or mirror it, before rendering.
    models: []
    cost: "-"
    inputs: A captured motion track and target cast.
    outputs: Group or symmetric choreography as data.
    risk: Pure data manipulation; free until it hits the generator.

  - id: controlled-hero-shot
    name: Controlled hero shot
    lane: [hero]
    stage: video
    tier: premium
    status: designed
    tags: [control, continuity-lock, pose-capture]
    summary: Pose, depth and reference conditioned video for a deliberately blocked hero shot.
    models: [Wan-VACE-class controllable video]
    cost: "≈A$0.30-1.50 per shot"
    inputs: Keyframes, a pose or depth map, cast and location assets.
    outputs: A directed shot with intended blocking and continuity.
    risk: The most demanding path; prove the shot at draft before rendering premium.

  - id: location-plate
    name: Location plate + lock
    lane: [release, hero]
    stage: scene
    tier: standard
    status: designed
    tags: [continuity-lock, asset]
    summary: Establish a consistent environment and reuse it across every shot in a scene.
    models: [Qwen-Image, SAM 2]
    cost: "generation cost only"
    inputs: A location description or plate image.
    outputs: A registered location asset shots can pull by name.
    risk: Lock it once; reuse buys both consistency and cheaper shots.

  - id: object-transfer
    name: Object / prop transfer
    lane: [release, hero]
    stage: scene
    tier: standard
    status: designed
    tags: [continuity-lock, control]
    summary: Detect, lift and place a consistent object or prop across shots.
    models: [SAM 2, Grounding DINO, inpaint]
    cost: "generation cost only"
    inputs: A reference of the object and the target frames.
    outputs: The prop held consistent through the scene.
    risk: Detection and matting run local; only the placement render costs GPU.

  - id: beat-montage
    name: Beat-synced lyric montage
    lane: [release]
    stage: assemble
    tier: standard
    status: planned
    tags: [beat-sync, assemble, multi-platform]
    summary: Cut approved clips to the beat and render for each platform. The forgiving, shareable lane.
    models: [librosa, ffmpeg]
    cost: "-"
    inputs: Approved clips and the song audio.
    outputs: A cut, rendered at 9:16, 1:1 and 16:9.
    risk: Local assembly; the cheapest way to a publishable, community-shareable piece.

  - id: micro-drama-cut
    name: Vertical micro-drama cut
    lane: [release, hero]
    stage: assemble
    tier: premium
    status: planned
    tags: [tts-lipsync, assemble, vertical]
    summary: A short vertical story with dialogue voice and lip-sync, cut for phone-first viewing.
    models: [Qwen3-TTS, LatentSync/InfiniteTalk, ffmpeg]
    cost: "voice + lipsync + assembly"
    inputs: A story seed, cast, and approved shots.
    outputs: A 9:16 micro-drama.
    risk: Lip-sync is a real gate; vet on a Recon pass first.

  - id: tts-voiceover
    name: TTS voiceover
    lane: [recon, release, hero]
    stage: voice
    tier: draft
    status: designed
    tags: [voice, narration, not-music]
    summary: Narration or spoken-word voiceover from a script. The non-music backbone.
    models: [Qwen3-TTS, Kokoro-82M]
    cost: "-"
    inputs: A script and a chosen voice.
    outputs: A voice track ready to cut against picture.
    risk: Runs light and local; Kokoro is CPU-viable for fast drafts.

  - id: voice-clone
    name: Voice clone (character / singer)
    lane: [recon, hero]
    stage: voice
    tier: standard
    status: designed
    tags: [voice, likeness, asset]
    summary: A cloned voice for a recurring character or singer, registered like a cast asset.
    models: [Qwen3-TTS, Chatterbox v3]
    cost: "-"
    inputs: A few seconds of consented reference audio.
    outputs: A reusable voice profile in the asset registry.
    risk: Keep consent and private voices local; mark the asset private.

  - id: stt-transcribe
    name: Speech-to-text transcribe
    lane: [recon, release, hero]
    stage: voice
    tier: draft
    status: designed
    tags: [stt, not-music, captions-source]
    summary: Transcribe dialogue or voiceover to timed text. Feeds captions, edits and search.
    models: [Whisper large v3, faster-whisper]
    cost: "-"
    inputs: Any audio or video with speech.
    outputs: A timed transcript (word and segment timestamps).
    risk: Local and cheap; distil/faster variants run near realtime.

  - id: autocaptions
    name: Auto captions / subtitles
    lane: [release, hero]
    stage: voice
    tier: draft
    status: designed
    tags: [captions, stt, assemble, multi-platform]
    summary: Generate SRT from transcription and burn styled captions for phone-first viewing.
    models: [Whisper large v3, ffmpeg]
    cost: "-"
    inputs: A clip with speech, or an existing transcript.
    outputs: An SRT and a caption-burned render.
    risk: Always proofread auto-captions before publish; names and lyrics trip them.

  - id: audio-dub
    name: Cultural-translation dub
    lane: [hero]
    stage: voice
    tier: standard
    status: planned
    tags: [voice, stt, translation, not-music]
    summary: Transcribe, translate, re-voice and re-sync into another language. For the cultural-translation work.
    models: [Whisper large v3, Qwen3-TTS, LatentSync]
    cost: "voice + lipsync cost"
    inputs: A finished clip and a target language.
    outputs: A dubbed, lip-synced version.
    risk: Translation needs a human check for meaning and register, not just literal words.