🅛🅣🅧 Add Video IC-LoRA Guide Advanced

Add Video IC-LoRA Guide Advanced is the same control-injection node as the regular Add Video IC-LoRA Guide, with one extra lever: it lets you dial how strongly a given guide pushes through self-attention, and optionally mask that influence to a region of the frame. You reach for it when you're stacking multiple control guides and need one to lead and another to defer, or when you want the control to apply to part of the frame and leave the rest free.

It's part of ComfyUI-LTXVideo, Lightricks' official pack. The base node covers "here's my depth/pose/canny video, follow it." The Advanced version is for the second-order problem: how much attention the model pays to this guide relative to everything else - and where.

How it works

Everything from the base node carries over - it VAE-encodes your control frames and injects them at a frame index, with a latent_downscale_factor to match the IC-LoRA's trained latent grid. The addition is attention_strength: a 0–1 control on how much this guide influences generation through the model's self-attention. Pair it with an attention_mask and the guide's influence is confined to the masked region, so you can, say, apply pose control to a character while leaving the background to the prompt.

The inputs that matter

• positive / negative / vae / latent / image - the same conditioning, encoder, latent, and control-frames set as the base node.
• frame_idx - where the guide begins (default 0).
• strength - overall guide strength (default 1).
• latent_downscale_factor - match it to your IC-LoRA (feed it from the loader's output).
• attention_strength - the new dial (default 1). Lower it to make this guide yield to others.
• attention_mask (optional) - restrict the guide's influence to a region.

There's also use_tiled_encode with tile_size / tile_overlap for encoding large control clips. Outputs are the modified positive, negative, and latent for your sampler.

Installing it

ComfyUI Manager → search LTXVideo → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo and restart. It's under Lightricks/IC-LoRA. The matching IC-LoRA goes in models/loras (from Lightricks' HuggingFace) and is applied via IC-LoRA Loader Model Only.

Where people get burned

• Using Advanced when the base node would do. If you're applying a single control over the whole frame at full strength, the plain Add Video IC-LoRA Guide is simpler and does the same thing. Reach for Advanced only when you need per-guide attention weighting or a spatial mask.
• Downscale mismatch. As with the base node, latent_downscale_factor must match the LoRA's trained grid - wire it from the loader rather than guessing.
• Mask/frame misalignment. The attention_mask has to correspond to the frames and resolution you're guiding; a mask built at the wrong size or for the wrong frames guides the wrong pixels. Mind LTX-2.3's divisible-by-32, (8n+1)-frame constraints.
• Turning every strength to max. With strength and attention_strength both at 1 on multiple stacked guides, they fight and the prompt loses. The point of this node is to balance guides - treat the two strengths as a mixer, not an on/off.