LTX2.5 Duration Head (CRT AutoDL)

The LTX2.5 Duration Head (CRT AutoDL) is the most niche node in the LTX2.5 stack, and it's worth knowing even if you won't use it every day. It loads a small model patch - ltx-2.5-duration-head-bf16.safetensors, auto-downloaded into ComfyUI/models/model_patches/ - and outputs a MODEL_PATCH that you apply to your diffusion model to steer variable-duration generation.

Think of it as the "length controller" of the LTX2.5 family. Video DiTs are trained at fixed-length latent templates, and their samplers generally expect a specific number of frames; the duration head is the piece that lets the model handle clips outside the default box - shorter or longer takes - with some awareness of where it is in the clip, rather than just being extended blindly. It's the kind of component that makes the "extend a scene smoothly" workflows possible instead of just chopping or padding frames.

What it is, mechanically

A MODEL_PATCH isn't a model - it's a small adjustment you fold into the main model, similar in spirit to a LoRA but targeting the model's runtime behavior rather than adding learned content. In the CRT stack it loads through ComfyUI's stock ModelPatchLoader machinery, so the wiring is standard: load it once, apply it alongside the LTX2.5 Model, and run your sampler. The output is specifically typed MODEL_PATCH, which tells you exactly where it's meant to plug in - no guessing about sockets.

There are no inputs to configure and one output. The auto-download pattern is the same as every AutoDL node: first run fetches the weights, and the bf16 patch is small (a few hundred MB), so this is one of the quick ones.

When you'd reach for it

If you're doing straight 5- or 10-second generations, you can leave this node out of the graph entirely and never miss it. It becomes relevant the moment you start pushing clip lengths outside the default and want the model to behave sensibly across the whole take - long-form, extension passes, or multi-scene work. Grab it when you're fighting frame-count artifacts; skip it when you're not.

Installing it

Standard CRT install:

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

or ComfyUI Manager → search CRT-Nodes, restart. First run downloads the patch.

Common issues

• MODEL_PATCH won't wire into the sampler - it's meant for the model-patch socket, not the model socket. If your sampler doesn't expose one, check you're on a recent ComfyUI; model patches are newer plumbing.
• No visible effect - the duration head is subtle by design; it matters on non-standard lengths, not on the default template. If you expected a dramatic change on a stock 10-second run, that's not what it does.
• Update breakage - red socket or NaN → right-click → Fix node (recreate).