# WhatDreamsCost-ComfyUI: a ComfyUI timeline editor with nothing pinned

> LTX Director gives ComfyUI users a timeline editor for LTX Video work, and the repository around it ships eight node files, three helper modules and no declared dependencies at all. The editor is the strong part, the packaging around it is the weak part.

**WhatDreamsCost/WhatDreamsCost-ComfyUI** — LTX Director and a variety of other custom ComfyUI nodes and workflows

- Repository: https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI
- Stars: 2,091 · Forks: 247
- Language: JavaScript
- License: GPL-3.0
- Published: 2026-10-14 · Updated: 2026-10-14 · Language: en
- Canonical page: https://hysenlabs.com/projects/whatdreamscost-whatdreamscost-comfyui

## A single clone lands a timeline editor and seven sibling node files

Installing a ComfyUI extension means putting the repository inside the nodes folder, and what lands there is a bundle rather than one tool. Eight top-level files are named after nodes: `ltx_director.py`, `ltx_director_guide.py`, `ltx_keyframer.py`, `ltx_sequencer.py`, `multi_image_loader.py`, `load_video_ui.py`, `load_audio_ui.py` and `speech_length_calculator.py`. Three more are plumbing rather than nodes, namely `__init__.py`, `patches.py` and `prompt_relay.py`, and a separate js directory holds the browser-side code. That split is worth pausing on, because the repository's primary language is recorded as JavaScript while every node entry point at the top level is Python. An `example_workflows/` directory ships ready-made graphs, which is the fastest way to see what any of this is for.

The package exists for one thing above all: LTX Director, a timeline editor for LTX Video work that the update log calls the author's most ambitious node. It is for people who cut a sequence in ComfyUI and want to trim, split, extend and re-roll parts of a clip without rebuilding a graph from scratch each time. Everything else in the folder is smaller work: file pickers for video and audio, an image loader, a keyframer and a speech length calculator.

## Installing takes one clone and two unpinned prerequisites

Setup is a single command. From your ComfyUI `custom_nodes` folder, run:

```bash
git clone https://github.com/WhatDreamscost/WhatDreamsCost-ComfyUI
```

One detail rewards copying rather than retyping: the owner segment of that URL is spelled WhatDreamscost, while the Repository field in `pyproject.toml` spells it WhatDreamsCost. The README documents no other install route.

There is no pip step and no requirements file. The `[project]` table in `pyproject.toml` carries four fields, the last of them `version = "2.0.5"`, and lists no dependencies. ComfyUI Manager is the other route, and the README points you at it if v2.0.0 has not appeared in its list yet, or tells you to fetch the nightly build instead. It also says ComfyUI-LTXVideo and ComfyUI-KJNodes must be updated alongside it, and states outright that the node cannot be used without updating ComfyUI-LTXVideo.

Here is what should make you stop. No packaged metadata backs those prerequisites. A ComfyUI version floor is present but commented out:

```toml
# "requires-comfyui" = ">=1.0.0"  # ComfyUI version compatibility
```

And the dependency list in the same `[tool.comfy]` block is empty:

```toml
includes = []
```

So both hard prerequisites live in README prose and in no file a tool could read. What that block does declare is publisher identity, with `PublisherId` set to whatdreamscost and `DisplayName` set to WhatDreamsCost-ComfyUI, plus a comment noting the file is used by Comfy Registry. Compatibility metadata that a registry could have checked is exactly what is absent.

## Three numbers describe the current build and none agree

The newest entry in the README's update section is v2.0.4. The install note still tells a reader to look for v2.0.0 in ComfyUI Manager. `pyproject.toml` declares 2.0.5. Nothing formal backs any of them: the repository publishes no GitHub releases, and the last push was on 2026-07-30.

That leaves the README update section as the only written record of what a given build changed, and the record is incomplete by its own admission. The v2.0.0 entry says the full list of changes would be added later, a promise the section never finishes, since it ends at v2.0.4 with nothing at all for 2.0.5. Manager visibility has been a problem before too. The v1.3.9 entry is one line about recent updates not showing in the manager, closed with the author's own admission that it took five tries to fix. An updater ends up reconciling three sources with no authoritative one among them.

## Prompt relay was bypassed on plain generations but not on timeline work

Every speed claim here attaches to one small helper. The v2.0.4 entry describes model caching so later runs reuse patched models instead of reloading them each time, framed as both a load-time saving and a cut in CPU overhead. The same entry says normal text-to-video, image-to-video and video-to-video generations no longer route through prompt relay and now run at the speed of other workflows. `prompt_relay.py` is one of only three helper modules at the top level, small enough to read end to end.

The catch is which paths got lifted. Only the plain generation paths stopped using relay, while timeline, keyframe and custom-audio work still does, so the gain is uneven across the workflows the node exists to support. The stated range is a potential 1.5-3x depending on hardware and parameters, wide enough that the caching detail tells you more than the headline. The same entry also lists improved `SageAttention` and `Triton` compatibility, which tells you what the runtime path involves without anyone measuring it for you.

## Timeline state is written to a json file the node manages itself

Timeline Saving and Loading writes the timeline and its settings to a json file, keeping every imported video, audio file and image plus each setting you changed. Retake Mode, added in v2.0.0, is labelled Beta and described as an early working experiment: you select a segment inside a shot and regenerate just that segment. v2.0.4 adds a filename output, a file dropdown, fixed drag and drop, and what it calls full Nodes 2.0 Support so the nodes render in both litegraph and the 2.0 frontend, alongside more thumbnails while scrubbing so you can trim more precisely.

Two gaps survive inside all that. The loader nodes are said to scan both the input folder and a whatdreamscost workspace folder, yet no path, no creation step and no setting for that folder appears in the README or among the top-level files. Meanwhile the widget layer was still being repaired through the 1.3.x line: v1.3.3 corrected the `duration_seconds` input and forced both duration widgets to stay visible, and v1.3.2 fixed `epsilon` overlapping `custom_width` plus invisible widgets when visibility was toggled from the settings menu. Scripting against those widget positions means coupling to a layout that moved several times.

## The stated successor still ships beside both of its predecessors

The v1.3.0 entry introduces LTX Director and LTX Director Guide and calls the pair the successor to LTX Sequencer and Multi Image Loader. Both predecessors remain in the same folder, so `ltx_sequencer.py` and `multi_image_loader.py` sit next to `ltx_director.py` and a fresh install exposes all three at once. You pick per graph, which is convenient until you are unsure which of the overlapping nodes a given workflow expects.

Attention has plainly moved on. The last word on Multi Image Loader is v1.2.9, a sweeping fix that dropped the drag-and-drop animations, added right-click copy, open and save on images and repaired the invisible hitbox under the node, with the author noting this time it was actually fixed. That same entry gave the Speech Length Calculator the text output it never had. Reference Image support in v2.0.4 is narrower still, riding the LTX Ingredients IC-LoRA track where images are dropped onto an IC-LoRA segment, with MSR and BFS recorded as the alternatives tried and passed over first.

## Where this bundle is the wrong tool to reach for

Several everyday jobs are served better by something simpler. If a graph is one text-to-video, image-to-video or video-to-video pass, the timeline adds nothing and the node still pulls two other packages into your update cycle. On an older ComfyUI build, the split between litegraph and Nodes 2.0 means you can meet the invisible widget and hitbox faults the log keeps repairing. And a frozen environment with pinned dependencies has nothing to pin against, since the version you get is decided by whatever nightlies those two companion packages happen to be on.

Support is one person, which shapes what you get when something breaks. The README links a Ko-fi page, the update log is written in the first person, and v1.0.0 through v2.0.0 are described as an author's own iteration. If you adopt this, budget for reading the update section before each ComfyUI bump rather than after it fails.

## Conclusion

Adopt it if you cut LTX Video timelines by hand and accept chasing two other packages whenever the node moves. Leave it for single-pass generation, where a timeline buys nothing, and for frozen environments that need resolved versions, since nothing in the package records one to resolve. Verify before you wire it into a shared graph: whether ComfyUI Manager lists the build you expect, whether ComfyUI-LTXVideo and ComfyUI-KJNodes are both current, and whether the whatdreamscost workspace folder your loaders scan actually exists on that machine.

## FAQ

### What is WhatDreamsCost-ComfyUI used for?

It is a collection of free ComfyUI custom nodes, led by LTX Director, a timeline editor for shaping LTX Video generations with prompts, keyframes, audio and imported clips. Smaller nodes cover video and audio loading, an image loader, keyframing and speech length calculation.

### How do I install WhatDreamsCost-ComfyUI?

Clone the repository into your ComfyUI `custom_nodes` folder, or install it through ComfyUI Manager. ComfyUI-LTXVideo and ComfyUI-KJNodes must be updated too, because the node cannot be used without an updated ComfyUI-LTXVideo.

### Which ComfyUI version does WhatDreamsCost-ComfyUI require?

No minimum is declared anywhere. The `requires-comfyui` line in `pyproject.toml` is commented out and the dependency list is an empty `includes = []`, so the only stated prerequisites are current copies of ComfyUI-LTXVideo and ComfyUI-KJNodes.

### What licence does WhatDreamsCost-ComfyUI use?

GPL-3.0, with the licence file referenced from the `[project]` table in `pyproject.toml`. Read that file yourself before deciding how the terms apply to what you build with it.

### Can I regenerate part of a clip with WhatDreamsCost-ComfyUI?

Yes, through Retake Mode, which selects a segment inside a shot and regenerates only that part. It is labelled Beta and described as an early working experiment, so treat output quality there as unsettled.

## Sources

- [Issues](https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI/issues)
- [License: GPL-3.0](https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI/blob/main/LICENSE)
- [README](https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI/blob/main/README.md)
- [WhatDreamsCost/WhatDreamsCost-ComfyUI on GitHub](https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/whatdreamscost-whatdreamscost-comfyui