A model-agnostic timeline node for ComfyUI, inspired by LTX Director's drag-and-drop shot editor — but the timeline UI itself doesn't know or care which video model you're using. Backend-specific adapter nodes translate the timeline into native conditioning calls for LTX, Wan, or Hunyuan.

Why adapters instead of one universal node

LTX, Wan, and Hunyuan condition video generation in genuinely different ways:

| Backend | Conditioning model | |---|---| | LTX | One continuous guide-latent stack; images injected at arbitrary frame indices via LTXVAddGuide | | Wan | Per-clip: WanImageToVideo (single start image) or WanFirstLastFrameToVideo (start+end pair), one conditioning+latent triple per clip | | Hunyuan | Per-clip: TextEncodeHunyuanVideo_ImageToVideo + EmptyHunyuanLatentVideo, similar shape to Wan |

A single node pretending to handle all three identically would either be shallow (lowest common denominator: text + one image, nothing else) or silently wrong for two of the three backends. Instead:

TV Director — the timeline UI. Build your shot list once: prompts, reference images, durations, transitions, camera hints, audio refs. Outputs a DIRECTOR_TIMELINE object (JSON under the hood) — a documented, versioned schema (see timeline_schema.py).
TV → LTX Adapter — flattens the timeline into one positive/negative conditioning pair + one guide latent, matching how LTX actually samples.
TV → Wan Adapter — returns a list of per-shot conditioning bundles (auto-pairs adjacent "first"/"last" shots into WanFirstLastFrameToVideo calls when both have images).
TV → Hunyuan Adapter — same per-shot list shape as Wan, using Hunyuan's native I2V/T2V encode nodes.
TV Wan Shot Iterator — pulls one shot's bundle out of a Wan or Hunyuan shot list by index, for feeding into a sampler in a per-shot loop.

Install

cd ComfyUI/custom_nodes/
git clone <your-repo-url> tensorvizion-director

Then install/update the backend node packs for whichever models you plan to use:

LTX: ComfyUI-LTXVideo (required for the LTX adapter — it calls LTXVConditioning / LTXVAddGuide directly rather than reimplementing them)
Wan: ComfyUI core, updated to a version with Wan 2.x support (WanImageToVideo / WanFirstLastFrameToVideo ship as built-in nodes)
Hunyuan: ComfyUI core with the classic HunyuanVideo node set (TextEncodeHunyuanVideo_ImageToVideo, EmptyHunyuanLatentVideo)

Each adapter checks for its required nodes at execution time and raises a clear error naming exactly what's missing — it won't fail silently or produce garbage output if a backend isn't installed.

Known limitations (read before reporting a bug)

Hunyuan 1.5 ships different node names than classic HunyuanVideo (e.g. HunyuanVideo15ImageToVideo). This adapter targets classic HunyuanVideo only; 1.5 support would need its own adapter variant.
Transitions (crossfade, hold, morph) are stored in the timeline schema as intent but none of the three adapters currently render them — all shots are conditioned independently and concatenated. Actual crossfade/morph blending between shots is a real feature gap, not an oversight to paper over.
Wan/Hunyuan multi-shot stitching is manual: you sample each shot separately and concatenate the resulting video clips yourself with a video-combine node. There's no automatic continuity-preserving stitch (e.g. carrying latent noise state between shots) yet.
Built and tested against node APIs current as of mid-2026; if LTXVAddGuide, WanImageToVideo, or the Hunyuan encode nodes change their signatures upstream, the adapters will need matching updates.

Timeline schema

See timeline_schema.py for the full, documented schema (SCHEMA_VERSION = 1). Key shape:

{
  "schema_version": 1,
  "global": { "fps": 24, "width": 768, "height": 512, "seed": -1, ... },
  "audio_tracks": [...],
  "shots": [
    {
      "id": "shot_abc123",
      "order": 0,
      "prompt": "...",
      "negative_prompt": "...",
      "duration_frames": 97,
      "image_ref": 0,
      "image_role": "first",
      "transition_in": "cut",
      "transition_out": "cut",
      "strength": 1.0,
      "camera_hint": "slow dolly in"
    }
  ]
}