LTX Director Guide (TJ 2.5)

LTX Director Guide (TJ 2.5) is the second half of the Director pair. The Director node is the editor - the timeline with its text, image and video segments. This node is the one that takes that plan and physically splices it into the latent and the conditioning so the sampler can see it.

You'll spend about ninety percent of your time in the Director and one percent here, but that one percent is where the render is actually steered.

What it's doing under the hood

The timeline hands over a guide_data payload: a list of guide images, the frame each one belongs at, and how strongly each should pull. The Guide node encodes each image through the video VAE at the latent's resolution, then uses ComfyUI's own LTX keyframe machinery (LTXVAddGuide) to append it into the latent stream at the right index - which is why the latent comes out longer than the clip and why the crop node exists downstream.

If you feed it motion_guide_data and pick a LoRA in ic_lora_name, it switches to IC-LoRA video guidance instead: the motion video is tiled and encoded as a reference, the LoRA is applied model-only, and each guide gets an attention entry on the conditioning so the reference actually influences the sampling. That's the path that lets you drive a shot's motion from a reference clip.

The inputs you'll actually touch

• positive / negative / latent / guide_data - all coming from the Director. Not much to decide.
• vae - the video VAE. Wrong VAE, wrong picture; this isn't the audio VAE.
• motion_guide_data + model + ic_lora_name - only if you're doing IC-LoRA motion guidance. Leave ic_lora_name on None and none of it engages.
• scale_by - latent scale factor, default 0.5 and the author recommends that for IC-LoRA. Worth knowing that this is a latent resize, not a UI zoom; weird numbers here show up as a blurry or chunky result, not an error.
• use_tiled_encode / tile_size / tile_overlap - turn it on when your guide images are big enough that encoding them in one go eats VRAM.
• image_attention_strength - how hard guide images pull attention, 0 to 1.
• msr_strength - only relevant to the MSR reference path, and the tooltip is worth reading: leave it at 0 for stage one, set ~0.4 on a refinement/upscale stage so references hold detail without repainting your opening. That's the fix for stage-2 mist and ghosting.

Outputs are positive, negative, latent, model, and latent_downscale_factor. Wire model into your sampler whenever you've loaded an IC-LoRA - the sampler must use the LoRA-patched model, not the pristine one. latent_downscale_factor is the matching scale for the IC-LoRA reference path; hand it to the stage that needs to know how far the references were downscaled.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE_STUDIO_ONE.git

Then restart and run the dependency script once - bash install_requirements.sh on Mac/Linux, install_requirements.bat on Windows. Two dependencies bite specifically here: the IC-LoRA motion path needs ComfyUI-LTXVideo (LTXAddVideoICLoRAGuide) in addition to the core LTXV nodes, and the module imports PyAV at load time, so a stripped-down ComfyUI install can fail to register the node at all. ComfyUI normally ships av already.

The gotcha that costs people an afternoon

If the timeline has image segments but the render ignores them, check that the Guide node is receiving the Director's guide_data output rather than a stale one, and that your guide images actually have a nonzero guide_strength - a segment with strength 0 is skipped silently, by design. The other one: guides are appended, never substituted. If you expect a first-frame image to replace frame 0 and instead weaves its way in, that's the model doing exactly what the mechanism says it does; drop the strength and it stops.