For the opt-in video.model_profile route, read PROFILED-VIDEO.md. It uses exact retained frame spans and full-master audio; legacy options are not automatically compatible, and no global LTX-2.5 default is promoted.

drama-video

A narrative-first sibling to music-video. The primary axis is dialogue — you author lines with pauses, generate them through ElevenLabs (along with an ambience bed), and cut shots against the per-line timestamps the API returns. Each shot is one LTX ia2v render (image + audio slice → video); shots chain via @last so the character's identity carries across cuts.

When to use this vs music-video

If the piece has both (song and dialogue over it), use music-video and add the dialogue as a per-scene audio: override.

Spec shape

title: "Fennie's Moment"
description: "Fennie practising alone late at night…"        # optional, for logs

character:
  name: Fennie
  anchor: ./path/to/character_anchor.png                     # primary identity image
  description: "Fennie — 17-year-old Korean ballet trainee…" # optional, available as {character} in prompts

reference_sheet: ./path/to/sheet.png                         # optional; for shots with `mode: ingredients` (see below)

video:
  fps: 24
  resolution: [1024, 576]
  negative: "cartoon, text, watermark, distorted face, still frame, low quality, extra fingers"
  tail_buffer_sec: 0.5                                       # same semantics as music-video

audio:
  # Either point at an already-generated scene audio mp3 + sidecar:
  file: ./outputs/scene_complete.mp3
  cues: ./outputs/scene_complete.cues.json                   # optional — if present, `drama_video.py plan` validates shot starts/ends against line cues
  # …or author inline and let this skill call the elevenlabs scripts:
  # spec: ./scene_audio.yaml                                 # an eleven_scene_audio.py spec; this skill renders it in stage 1

gate_confirm_audio: false
gate_confirm_anchors: true

shots:
  - label: opening_focus
    start_sec: 0.00
    duration_sec: 13.43                                      # ≤ 15s (LTX cap)
    image: "@anchor"                                         # @anchor | @last | path
    camera_lora: static                                      # optional
    camera_lora_strength: 0.8
    prompt: "Close-mid shot of {character} in a late-night dance studio…"
    anchor:                                                  # optional flux2 pre-gen anchor (same block as music-video)
      type: i2i
      reference: ./path/to/character_anchor.png
      prompt: "Wide shot of the studio, overhead spot, wooden floor…"

  - label: frustration
    start_sec: 13.43
    duration_sec: 10.85
    image: "@last"
    camera_lora: dolly-in
    prompt: "{character} sinks onto the floor…"

Shot durations

Hard cap 15 s per shot (same as music-video, for the same LTX-2.3 OOM reasons). If a cue range is longer, split it into two shots with the same prompt trend — you pick the visual cut point.

Continuation between shots (recommended for most drama)

A regular image: "@last" shot starts from the previous shot's last single frame — LTX re-interprets that one frame and often produces a visible cut even with the transition LoRA. For truly seamless continuation, use the continue_from_prev flag:

shots:
  - label: opening_focus          # shot 1 — uses anchor (scene setup)
    start_sec: 0.0
    duration_sec: 4.68
    image: "@anchor"
    anchor:
      type: i2i
      reference: ./media/character.png
      prompt: "The character in the studio, overhead light, …"
    prompt: "Close-mid shot, she exhales, rises onto her toes …"

  - label: first_attempt          # shot 2 — continues from shot 1
    start_sec: 4.68
    duration_sec: 7.96
    continue_from_prev: true      # ← locks first 1s to shot 1's last 1s
    overlap_seconds: 1.0          # ← default 1.0 (0.5–1.5s is the sweet spot)
    overlap_strength: 1.0         # ← default 1.0 (hard lock); 0.85 for softer
    prompt: "She attempts the spin, mid-motion, arms extended …"
    # no `image:` or `anchor:` needed — LTX picks up from shot 1's tail

How it works: the pipeline invokes the continuation comfyui workflow, which loads the previous shot's mp4, extracts the last N frames via LoadVideo → GetVideoComponents → GetImageRangeFromBatch, and injects them as a multi-frame LTXVAddGuide(frame_idx=0, strength=overlap_strength) on an empty latent. The rest of the shot is generated from the new prompt + audio slice.

When to use continue_from_prev: same physical space, time flows forward, character continues to be present. Most drama shots fit this.

When to fall back to anchor + fresh ia2v: hard scene change (different room, different character, flashback). Skip continue_from_prev on that shot; give it a fresh anchor: block.

Ingredients mode (silent inserts)

For B-roll / establishing / no-dialogue shots that must keep the recurring cast, props and locations consistent, set mode: ingredients on the shot and provide a project reference_sheet: (black-bg element panels — build it with storyboard/scripts/generate_reference_sheet.py, aspect matching video.resolution). The shot renders via the LTX-2.3 reference-sheet IC-LoRA:

reference_sheet: ./sheet.png
shots:
  - label: pier_insert
    start_sec: 41.0
    duration_sec: 5                     # ~121f @ 24fps trained bucket
    prompt: "wide insert: Marco alone on the pier at dusk, holding the trident"
    mode: ingredients

Silent by design — ingredients shots carry no lipsync/dialogue (identity comes from the sheet, not the identity_anchor); the full audio track is muxed over the timeline at assemble. Use them only where on-screen speech isn't needed. mode: ingredients is checked first in the dispatch (over continuation/ia2v). Per-shot reference_sheet, negative, ingredients_lora_strength (def 1.4) and ingredients_reference_strength (def 1.0) override the defaults. See the comfyui skill's ingredients command.

CLI

drama_video.py plan      <spec.yaml>           # show shots + audio alignment
drama_video.py audio     <spec.yaml>           # render audio via elevenlabs (if audio.spec set)
drama_video.py anchors   <spec.yaml>           # pre-render all shot anchors via flux2
drama_video.py shot N    <spec.yaml>           # render one shot (1-indexed)
drama_video.py shots     <spec.yaml>           # render all shots serially (skips existing)
drama_video.py assemble  <spec.yaml>           # vconcat + overlay the full audio
drama_video.py all       <spec.yaml>           # 1→5 end-to-end, with quality gates
drama_video.py status    <spec.yaml>           # show what's on disk

Each stage is idempotent — mp4s and anchors are skipped if already present. To force a re-render, rm the artifact and re-run the stage.

Output layout

<project>/
  spec.yaml
  audio/                       # if audio.spec was used
    scene_complete.mp3
    scene_complete.cues.json
  shots/
    001-opening_focus-anchor.png
    001-opening_focus.mp3      # audio slice for this shot
    001-opening_focus.mp4
    001-opening_focus-last.png
    …
  final.mp4
  run.log

Prompt conventions

• The top-level character.description is available as {character} inside any shot prompt (and anchor prompt). Recurring-character substitution just like music-video's subjects: block — useful so you don't copy/paste the identity description into every shot.
• Inside anchor prompts, refer to the character by name or via {character} so flux2 i2i preserves identity when blending with a new scene environment.
• Keep shot prompts tight: one sentence for setting, one for action, one for camera motion. Long shot prompts compete with the character anchor and drift identity.

Gotchas

• Shot 1's @anchor resolves to character.anchor; later shots with image: "@last" chain from the prev shot's -last.png. First shot must have either @anchor or a literal image: path.
• The audio block must exist before shots stage runs. If you set audio.spec:, the audio stage renders it; otherwise audio.file: must already exist on disk.
• Shot start_sec + duration_sec must stay within the audio file's length. plan checks this.
• If your audio has cues.json, plan warns when a shot's start_sec lands mid-line (cuts across a word).