LTX 2.3 12GB GGUF Director v2 ComfyUI GGUF Workflow and One Click Installer

LTX 2.3 Director 2 all in one ComfyUI one click Windows installer — a single setup that installs ComfyUI Portable, every custom node, and every model needed to run the LTX Director 2 timeline node on a 12 GB card using GGUF quantized models. No manual model hunting, no folder placement, no chasing down which node pack broke your graph.

LTX Director 2 turns ComfyUI into an actual video editor. Instead of maintaining separate workflows for text to video, image to video, image + audio to video, and multi keyframe sequences, you get one timeline node where you drop in images, layer prompts across segments, add custom audio, and now edit existing video directly — trim, split, combine, and extend clips with prompts and keyframes without ever leaving the node. Version 2 also adds IC-LoRA support, so you can drag a video onto the IC-LoRA track for structural control, VFX, and restoration passes. The catch is that getting there normally means a portable ComfyUI install, a pinned CUDA build of PyTorch, six custom node repos that all need to be on current versions, and seven model files scattered across four different Hugging Face accounts. This installer does all of it from one double click.

The free LTX 2.3 Director 2 GGUF workflow is available on my Patreon and is included with this installer for convenience. If you want the full walkthrough of how the Director node works, read that post first — the links are below.

What's Included in This LTX 2.3 Director 2 Installer

• ComfyUI Windows Portable — downloaded, extracted, and configured, with the CUDA 12.8 build of PyTorch installed so LTX 2.3 runs on RTX 30, 40, and 50 series cards without the usual torch version mismatch.

• LTX Director 2 timeline node — the full WhatDreamsCost node pack, plus every dependency it requires. Director 2 will not load against outdated ComfyUI-LTXVideo or KJNodes, which is the single most common reason people can't get it running.

• LTX 2.3 22B Distilled GGUF workflow — the low VRAM GGUF build of the 22B distilled model with a two stage generate and spatial upscale pass, tuned to fit inside 12 GB.

• Optional automated model downloads — the installer asks whether to pull the models at the end, so you can skip the download if you already have them locally.

Additional Resources

LTX Director node by WhatDreamsCost - https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI

Youtube Tutorial on how to use LTX Director - https://www.youtube.com/watch?v=vM60pJJqqEI

Preloaded Models Within the Installer (Low VRAM)

Every model below is downloaded automatically and placed in the correct folder by the installer. The paths are listed for anyone setting up manually.

• ltx-2.3-22b-distilled-1.1-Q3_K_S.gguf (place in ComfyUI/models/unet/) – Hugging Face Link

• gemma-3-12b-it-UD-Q4_K_XL.gguf (place in ComfyUI/models/clip/) – Hugging Face Link

• ltx-2.3_text_projection_bf16.safetensors (place in ComfyUI/models/clip/) – Hugging Face Link

• LTX23_video_vae_bf16.safetensors (place in ComfyUI/models/vae/) – Hugging Face Link

• LTX23_audio_vae_bf16.safetensors (place in ComfyUI/models/vae/) – Hugging Face Link

• taeltx2_3.safetensors (place in ComfyUI/models/vae/) – Hugging Face Link

• ltx-2.3-spatial-upscaler-x2-1.1.safetensors (place in ComfyUI/models/latent_upscale_models/) – Hugging Face Link

This build uses the GGUF quantized text encoder and the Q3_K_S quantized 22B distilled model to keep the workflow inside 12 GB of VRAM. The installer omits the standard LTX 2.3 22B checkpoints (BF16, FP8) for the same reason, but if you have the headroom they drop straight into the same setup and can be downloaded from Kijai's repository: LTX-2.3 22B Diffusion Models

Custom Nodes w/ commands needed for Portable Windows package

All custom nodes required by the Director 2 workflow are cloned and have their requirements installed automatically, including ComfyUI-GGUF, KJNodes, ComfyUI-LTXVideo, and the WhatDreamsCost pack itself. Nothing here needs to be cloned by hand.

• ComfyUI Manager - https://github.com/ltdrdata/ComfyUI-Manager

Command to install requirements:

.\python_embeded\python.exe -m pip install -r .\ComfyUI\custom_nodes\ComfyUI-Manager\requirements.txt

Use the ComfyUI Manager's Install missing custom nodes feature to install any additional nodes that may not work initially after installation.

System Requirements

• NVIDIA RTX 30XX / 40XX / 50XX GPU

• CUDA-compatible GPU (minimum 12 GB VRAM)

• At least 40 GB free storage

• Windows 10 or 11

Usage Notes

• Place the install .bat file in its own separate directory dedicated to this project, then double click it to start the installation.

• The installer will pause partway through and ask whether you want to download the models. Press Y to download them, or N if you already have them and would rather move them in manually. The window waits for your answer, so it will not time out.

• Load the provided workflow into ComfyUI.

• Verify all model checkpoints are mapped correctly in your Loader nodes.

• Director 2 requires current versions of ComfyUI-LTXVideo and ComfyUI-KJNodes. If you later update the Director node through ComfyUI Manager, update those two as well or the node will fail to load.

• Build your generation on the timeline first, then hit Run. Images go on the main track, prompts attach to individual segments, and the global prompt applies across the whole timeline — a segment prompt will override it for that stretch.

• For IC-LoRA work, drag your reference video onto the IC-LoRA track rather than wiring a separate conditioning chain. It stays compatible with prompt relay, keyframes, and custom audio.

• The workflow runs a two stage pass — a base generation followed by a spatial upscale. If you're tight on VRAM or just testing a prompt, bypass Stage 2 to iterate faster, then re-enable it for the final render.

• Keep your first tests short. Timeline duration is the main driver of VRAM and generation time here, so get the composition right at a few seconds before extending it out.

• If you get black frames on an audio segment, check that your audio actually spans that part of the timeline. Gaps in the audio track are handled differently from silence.