AD_LTX_IC_lora

If plain text-to-video LTX keeps giving you the wrong pose, wrong composition, or wrong structure, this is the node that fixes it - it applies an IC-LoRA (In-Context LoRA) against a guide image, which is Lightricks' official mechanism for structural control on LTX-13B and later. IC-LoRAs take a control video/image (pose, depth, or canny edges) and condition generation on its structure rather than leaving composition entirely up to the prompt. Apt_Preset wraps that as a node that slots in between your config loader and your sampler.

This matters more than it sounds like: LTX's biggest acknowledged weakness against Wan is prompt adherence on complex scenes. An IC-LoRA sidesteps that problem entirely for anything you can express as a reference image instead of words - if you can show it the pose, you don't have to describe the pose.

How it works

It takes your model and conditioning, plus a lora_name (the actual IC-LoRA weights - pose, depth, or canny) and a guide image, and patches the model so generation tracks that guide's structure. Because LTX videos can run longer than what a single guide frame can cleanly condition, this node also handles splitting a long generation into segments, each anchored to its own guide frame.

The inputs and outputs that matter

• model, positive, negative, img_latent, img_vae - the standard chain from your LTX config.
• image - your guide/control image.
• lora_name - which IC-LoRA to apply (pose, depth, canny, or whatever else Lightricks or the community has shipped).
• strength_model (default 1, range -100 to 100) - standard LoRA strength.
• frame_idx (default 0, range 0–9999) - which frame position the guide image maps to.
• guide_strength (default 1, range 0–1) - how strongly the guide's structure is enforced versus letting the model deviate.
• guide_preset (default auto_balanced) - a preset combo for how the guide is applied; worth trying the default first before hand-tuning guide_strength/max_segments yourself.
• max_segments (default 4, range 1–12) and max_guide_frames (default 17, range 1–257) - how a long generation gets chopped into guide-anchored chunks.
• lowGpu_mode (default false) - trade speed for lower VRAM use.

Outputs: model, positive, negative, img_latent, img_vae - feed these straight into AD_LTX_sampler in place of the un-patched versions.

How to install it

Via ComfyUI Manager: search "ComfyUI-Apt_Preset". Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git

Then install.bat (Windows) or pip install -r requirements.txt (Linux/Mac), and restart. You'll also need the actual IC-LoRA weight files - Lightricks ships official pose/depth/canny IC-LoRAs on HuggingFace for the LTX line - downloaded and placed wherever your ComfyUI LoRA folder is, so they show up in the lora_name dropdown.

Common issues & troubleshooting

lora_name dropdown is empty or missing the LoRA you want. You need the IC-LoRA .safetensors file physically in your LoRA folder before ComfyUI starts - this node doesn't download anything for you.

Guide isn't being followed strongly enough. Push guide_strength up toward 1, or check that frame_idx actually lines up with where in the sequence you want the guide to anchor.

Long generation looks disjointed between segments. Try adjusting max_segments/max_guide_frames - if your guide content changes significantly over the video's length, more, shorter segments (each closer to its own guide) will track better than a few long ones.

Running low on VRAM with a guide LoRA plus a long clip. Flip on lowGpu_mode before dropping resolution or frame count - structural guidance adds real overhead on top of a normal LTX pass.