A typical LTX video workflow is 100 to 270 nodes. This one is 35. That is not because it does less — it is because the orchestration moved into a director node that takes a timeline instead of a graph.
Whether that is an improvement depends on what you are doing, and the trade-off is worth understanding before adopting it.
What the graph actually loads
| Node | Value |
|---|---|
CheckpointLoaderSimple |
ltx-2.3-22b-dev-dare-ties-distilled-1.1.safetensors |
DualCLIPLoader |
gemma_3_12B_it_fp8_e4m3fn.safetensors + ltx-2.3_text_projection_bf16.safetensors
|
VAELoaderKJ |
LTX23_audio_vae_bf16.safetensors, main_device, bf16
|
LoraLoaderModelOnly |
OmniNFT-comfyui.safetensors at 1.0
|
KSamplerSelect |
euler |
Two things stand out.
The dual CLIP setup. LTX 2.3 uses Gemma 3 12B as its text encoder, plus a separate text projection file. That is a larger language model than most image pipelines use, and it is why LTX handles longer, more specific prompts than CLIP-based video models.
VAELoaderKJ with main_device. The KJ loader variant explicitly places the audio VAE on the main device. On a multi-GPU or CPU-offload setup that placement decision is not automatic, and audio VAE on the wrong device is a common source of silent failure.
euler rather than euler_ancestral_cfg_pp. The plainer sampler, which matches the fact that a director node is driving structure rather than exploring.
What "one timeline" buys
In a conventional LTX graph, arranging images, text and audio means wiring conditioning paths by hand. Every added element is more links, and every link is a place to get the ordering wrong.
In the director node, those become tracks. The ordering is explicit, editable, and visible. For content where the sequence is the point — a narrated sequence, a music-driven piece, a multi-shot story — that is a real reduction in error surface.
What it costs
Debuggability. When a 250-node graph produces something wrong, you can trace the links. When a director node produces something wrong, the logic is inside the node and you are reading its parameters instead of its wiring.
Reuse. A 35-node director graph is not a component you drop into another pipeline. Every other LTX graph in this family expects conditioning links; the director expects a timeline. They do not compose.
Fine control. The director covers the orchestration layer, not the sampling layer. Per-stage sigma, anchor behaviour and block scheduling live outside it.
Where this is the right tool
- Content where the sequence is the creative decision, not the sampling settings.
- Handing a workflow to someone who should not have to understand LTX's conditioning graph.
- Iterating on structure — how many beats, in what order — where the conventional approach means rewiring each time.
Where it is the wrong tool
- You need per-stage sampling control.
- You are building a component to reuse inside a larger graph.
- You are debugging why a specific conditioning path is behaving strangely.
For those cases, use the explicit graphs. The director node compresses well but it compresses the debugging surface too.
Workflows
One graph, published on RunningHub with invite code rh-v1111: the LTX director node workflow.
Files
The ComfyUI workflow export is in the companion repository:
https://huggingface.co/AI-KSK/ksk-video-243-comfyui-workflows
Video walkthrough (Chinese): https://youtu.be/4-W-j7KNLTA
Bilibili: https://www.bilibili.com/video/BV1FhLz6JEgJ/
Notes
- Workflow file only. Model weights are not included; each model keeps its own licence.
- This is a tutorial export, not a packaged product.
- Node count and loader values quoted above are read from the published JSON file.
- Independent AI-KSK release. Not affiliated with or endorsed by Lightricks or RunningHub.
Optional API platform - 88API (affiliate link)
Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.
https://88api.ai/sign-up?aff=usx3
Disclosure: I may earn a commission from eligible purchases after you register through this referral link.
Top comments (0)