LTX Director in ComfyUI: Detailed LTX 2.3 Timeline, Prompt Relay, Keyframe, and Audio Guide

LTX Director is a ComfyUI custom node that turns LTX 2.3 video generation into a timeline-driven directing workflow. Instead of writing one large prompt and hoping events happen in the right order, you create time segments, drop images or audio onto the timeline, assign prompts to specific spans, and use guide images as keyframes that the model tries to reach.

1. Confirm ComfyUI works first. Open your ComfyUI instance and run any known-good workflow before adding new custom nodes.
2. Install WhatDreamsCost-ComfyUI. Use ComfyUI Manager and search for What Dreams Cost, or clone the GitHub repo into ComfyUI/custom_nodes/.
3. Choose the nightly/latest build if needed. The tutorial and README both emphasize using the latest/nightly version when the standard Manager listing lags behind.
4. Update dependencies. Update ComfyUI-LTXVideo and ComfyUI-KJNodes. The repo README states ComfyUI-LTXVideo must be updated for LTX Director to work.
5. Restart ComfyUI. Custom node changes usually require a restart or full refresh.
6. Load an LTX Director workflow. In ComfyUI templates, search LTX Director, or import one of the example workflows from the GitHub repo.
7. Run a conservative test. Use a short duration, low resolution, and one simple prompt before building a complex scene.

The node is designed to behave like a compact video editor inside ComfyUI. These are the controls the tutorial calls out.

1. Load a starting image. Drag an image onto the image track. The tutorial demonstrates with an image made in Z-Image, but any suitable source image can be used.
2. Add one simple prompt. Keep the first prompt easy to evaluate. Example: the woman raises her hand and waves it.
3. Lower the resolution for the first run. This makes setup errors cheaper and faster to diagnose.
4. Run the workflow. Confirm that the image influences the output and that the prompted motion appears at least generally.
5. Only then add complexity. Once the basic path works, add more segments, audio, or keyframes.

Normal LTX prompting can describe a sequence, but it may not perform each event at the exact moment you intended. Prompt Relay solves that by letting you distribute instructions across the timeline.

Segment 1: the woman stands in the rain, looking toward the camera
Segment 2: lightning strikes in the background, lighting her face
Segment 3: the woman walks toward the camera and says in a sad voice, "I'm sorry"

The tutorial highlights two major reasons outputs fail to follow prompts.

Weak: action scene with robot
Better: a heavy robot bends down, grips the small car with both hands, lifts it slowly, then turns toward the glass office building

Too much: the robot picks up the car and throws it at the building in 2 seconds
Better: Segment 1 lifts car; Segment 2 turns body; Segment 3 throws car

In LTX Director, every image on the timeline is a guide image. That means it can be used as a keyframe at the beginning, middle, or end of a shot. You are not limited to classic image-to-video where the image must be the first frame.

1. Start with the frame you care about. Example: a hero walking toward camera while an explosion fills the background.
2. Place that image where the moment should happen. If the explosion should occur at second 4, slide the image later on the timeline.
3. Prompt what happens before it. The model is guided toward the keyframe, so the preceding segment can describe the build-up.
4. Add more images for a longer shot. The tutorial shows that multiple guide images can flow together into a longer camera move.
5. Adjust guide strength. Increase it when the image is not respected enough; reduce it if the output feels too rigid or cannot move naturally.

LTX Director can import your own audio and use it to drive the video. This is useful for dialogue, reactions, or physical movement that should align with a sound cue.

1. Drag or load the audio clip onto the audio track.
2. Trim it to the part you need. Keep the first test short and audible.
3. Toggle custom audio on. If this setting is off, the workflow may ignore the imported audio path.
4. Write a prompt that explicitly describes speech. Avoid only writing person speaks.
5. If lip sync fails, include the actual words. The tutorial says typing what is being said can make the lip-sync behavior much more reliable.

Better lip-sync prompt pattern:
A close-up of the tired man speaking clearly, his mouth forming the words: "Here I am, old and tired and fat." He looks exhausted and pats his stomach on the final phrase.

Global prompt examples:
continuous heavy rain, wet pavement reflections, handheld documentary camera
constant red emergency alarm light pulses in the background
wind blows through the character's hair throughout the entire shot

• The LTX Director node appears in ComfyUI and the example workflows load without missing-node errors.
• Duration seconds and frames update together.
• Dragging an image or audio file onto the timeline creates the expected media item.
• A low-resolution image-to-video test completes without crashing.
• Changing custom width/height changes the output size.
• A later prompt segment changes what happens later in the clip.
• Moving a guide image later in the timeline changes the generated build-up to that frame.
• Custom audio only affects generation after the custom-audio toggle is enabled.

• What Dreams Cost — How to use LTX Director — primary tutorial transcript and chapter structure.
• WhatDreamsCost-ComfyUI GitHub repo — install method, current version notes, dependency warnings, README, and issue link.
• LTX Director example workflows — full and subgraph workflow variants for importing manually.
• Official LTX 2.3 prompting guide — recommended by the video for LTX-specific prompt structure and common mistakes.
• Source notes for this guide — retrieval details and transcript-derived procedure notes.