技能创建器
skill-creator
创建、改进和评估自定义技能,把专业流程转化为可重复使用的 AI 能力。
ltx-director【数据与研究】Drive the LTX Director (Timeline) node. Its Add Image/Text/Audio buttons are DOM-only and cannot be clicked by an agent; edit the hidden timeline_data JSON widget instead. Load when a workflow contains LTXDirector / LTXDirectorGuide / PromptRelayEncodeTimeline, or when asked to add, move, retime, or remove timeline segments (image / text / audio / mo…
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
LTXDirector (pack: WhatDreamsCost-ComfyUI, category WhatDreamsCost) is a
video timeline editor node. Its on-canvas UI has Add Image, Add Text
and Add Audio buttons.
They are not node inputs. The pack builds them as raw DOM elements with JS handlers:
addTextBtn.addEventListener("click", () => this.addTextSegmentFreeSpace());
Panel/MCP tools drive the LiteGraph node model (widgets + inputs). They cannot invoke arbitrary DOM handlers, so "click Add Text" is impossible. This is a hard limitation, not flakiness. Say so plainly rather than retrying.
timeline_dataEverything those buttons do is serialized into ONE string widget, timeline_data.
The pack's own code treats it as "the absolute source of truth". Set it and the
editor renders it.
It is listed in the pack's HIDDEN_WIDGET_NAMES, so it does not appear in the
node's visible widget list, but it is an ordinary input and is settable:
create_workflow(action="modify", workflow=<the graph>, operations=[
{ op: "set_input", node_id: "42",
input_name: "timeline_data", value: "<json string>" },
])
The value is a JSON string, not an object. Empty state is "{}".
Sibling hidden widgets: local_prompts, segment_lengths, guide_strength,
audio_data, use_custom_audio, inpaint_audio, use_custom_motion,
override_audio.
{
// TRACK GATES — a track's segments are IGNORED unless its gate is true.
"mainTrackEnabled": true,
"audioTrackEnabled": false,
"motionTrackEnabled": false,
// must mirror the node's start_frame / duration_frames widgets (see Edges)
"normalStartFrame": 0,
"normalDurationFrames": 4393,
"global_prompt": "",
"overrideAudio": false,
"inpaint_audio": false,
// retake mode
"retakeMode": false, "retakeStart": 24, "retakeLength": 48,
"retakePrompt": "", "retakeStrength": 1, "retakeVideo": null,
"retake_global_prompt": "",
// display-only
"propHeight": 90, "globalPropHeight": 60, "showFilenames": true,
"segments": [], // main track: image AND text segments
"motionSegments": [],
"audioSegments": []
}
segments){ "id": "seg1", "start": 0, "length": 48, "prompt": "wide shot, neon city", "type": "text" }
segments){
"id": "1784657553220x4fxu",
"start": 0,
"length": 241.68,
"prompt": "",
"type": "image",
"imageFile": "whatdreamscost/Nintendo_ZeldaUrbosaHotspringKF_00002_.png",
"imageB64": "/api/view?filename=Nintendo_ZeldaUrbosaHotspringKF_00002_.png&type=input&subfolder=whatdreamscost"
}
imageB64 is a misnomer. It holds a /api/view URL, not base64. So an
image segment points at a file already in ComfyUI's input dir. Full
agent-drivable recipe:
upload_image (action:"image") → puts the file in the input dir (note its subfolder/name)imageFile = "<subfolder>/<name>.png"imageB64 = "/api/view?filename=<name>.png&type=input&subfolder=<subfolder>"audioSegments){
"id": "17846562234967gjvk", "type": "audio",
"start": 0, "length": 4393, "trimStart": 0, "audioDurationFrames": 4393,
"audioFile": "whatdreamscost/Intergalactic Hip Hop.mp3",
"fileName": "Intergalactic Hip Hop.mp3",
"waveformPeaks": [0.047, 0.274, 0.537, "…~200 floats…"]
}
waveformPeaks is the rendered waveform. It is cosmetic. Omit it or pass [] if
you're writing a segment programmatically; the audio still plays. Don't fabricate
plausible-looking peaks and imply they were measured.
audioSegments does nothing while
audioTrackEnabled is false. Set the gate in the same edit. (The reference
workflow ships an audio segment with the track OFF, which is easy to misread
as broken.)normalDurationFrames ==
duration_frames widget == segment_lengths widget (observed: all 4393),
and normalStartFrame == start_frame. Change the timeline length and you
must update all of them.start / length are pixel-space frames and may be
fractional (241.68). duration_seconds / frame_rate / time_units are
display concerns.global_prompt exists twice, inside timeline_data and as a forceInput
socket on the node. In the reference workflow the socket is wired from a
PrimitiveStringMultiline; prefer the wired source and keep them consistent.guide_data must go somewhere. LTXDirector.guide_data →
LTXDirectorGuide. That partner node is what injects the keyframes.Power Lora Loader (rgthree) ─ model ─┐
DualCLIPLoader ──────────── clip ────┤
VAELoaderKJ ───────────── audio_vae ─┤ LTXDirector #42
PrimitiveStringMultiline ─ global_prompt ─┘
│ model → LTX2_NAG
│ positive → ConditioningZeroOut, LTXVConditioning
│ audio_latent → LTXVConcatAVLatent
│ guide_data → LTXDirectorGuide ← required partner
│ frame_rate → AudioToFrameCount, VHS_VideoCombine
└ combined_audio→ AudioToFrameCount, PreviewAudio, …
get_workflow (action:"query") with types: ["LTXDirector"], fields: "detail"
(or panel_query_graph on the live canvas).timeline_data and JSON.parse it. Never hand-splice the string.normalStartFrame / normalDurationFrames in sync with the widgets (edge 2).set_input.create_workflow (action:"validate").PromptRelayEncodeTimeline (pack: ComfyUI-PromptRelay) carries the same
timeline_data hidden-widget pattern, so this skill's approach applies there too.
Any node whose controls are DOM widgets over hidden state is invisible to clicking and must be driven through its state widget. LTX Director is the clearest example.