LTX US Models Pipe (CRT)

LTX 2.3's unified sampler is a many-input beast: base model, video VAE, audio VAE, CLIP, plus optional mode-specific models for depth control, outpainting, upscaling, a latent upscaler, and a depth-estimation model. Wire all that directly into the sampler and the graph looks like a plate of spaghetti. LTX 2.3 US Models Pipe (CRT) is the node that gathers every model into one bundle so the unified sampler takes a single tidy input instead.

It's a pure plumbing node - no computation happens here. It collects what you give it, wraps it in an LTX23_US_MODELS_PIPE structure, and hands the whole thing to CRT_LTX23UnifiedSampler (the "US" in the name). If you're following any of the CRT pack's LTX 2.3 workflows, this is the node sitting between your loaders and the sampler, and its only job is to not drop anything.

The inputs

Three are required, and they're the non-negotiables of any LTX 2.3 run:

• model - the transformer. This is where the AutoDL loaders plug in (fp8, NVFP4, or GGUF Q5).
• vae - the video VAE.
• audio_vae - LTX 2.3 is one of the few video models with its own audio path, and it needs its own VAE for that.
• clip - the text encoder side (Gemma + projection, via the pack's CLIP nodes).

Then the optional inputs, which are where the interesting LTX 2.3 modes come from:

• model_union_control - a model merged with the Union Control / IC-LoRA. The tooltip calls this "V2V Depth Control" - feed the model that's already got the union-control LoRA applied here.
• model_outpaint - the model merged with the IC-LoRA Outpaint (that's what the AutoDL IC Outpaint LoRA node produces).
• model_upscale - the model merged with the upscale IC-LoRA.
• spatial_upscale_model - an optional latent upscaler for HQ generation and V2V upscale mode.
• da3_model - a Depth Anything V3 model, used to build V2V depth guides when no precomputed depth is connected.
• latent_downscale_factor - the compression factor the upscaler LoRA reports. The tooltip is blunt: connect the loader output rather than guessing this value. The IC-LoRAs carry it in their metadata and the pack reads it out, so don't hand-type it.

Output is a single models_pipe that goes straight into the unified sampler's models_pipe input.

How to use it

The intended flow: AutoDL loaders for the base model, VAEs, and CLIP → this node bundles them → pick your mode (depth/outpaint/upscale) by connecting the matching optional models → unified sampler. If you're doing a plain text-to-video, the required four inputs are all you need and the optional slots stay empty - the pipe happily passes through with only the base stack.

Install and gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

or ComfyUI Manager → search "CRT-Nodes". It's part of the LTX 2.3 subsystem, so it's only registered when the unified sampler module imports cleanly. The gotchas are all about matching: the optional models must already have their LoRAs merged before they enter the pipe (this node doesn't merge anything), and latent_downscale_factor really should come from the loader, not your judgment - wrong values here silently shift how the sampler interprets latents. If your depth or upscale mode produces garbage, nine times out of ten the pipe is fine and the mismatch is upstream in which model you merged with which LoRA.