LTX Attention Override

This is a small selector node, and it lives in the "ltxtricks" corner of the pack - the set of nodes ported from Logtd's community ComfyUI-LTXTricks project, which brought RF-inversion editing and attention manipulation to LTX. LTXAttnOverride doesn't do anything visible on its own. Its whole job is to hand the other trick nodes a list of which transformer attention layers to operate on. Think of it as a targeting config, not an effect.

Why would you target specific layers? Because in a diffusion transformer, different attention blocks carry different responsibilities - some hold coarse structure and motion, others fine texture. Techniques like skip-layer guidance and attention injection get their power (and their side effects) from touching the right blocks. Rather than bake a fixed choice into every node, the ltxtricks design factors that choice out into this one node so you can experiment.

How it works

You type a set of layer identifiers into a text box; the node parses them and emits an ATTN_OVERRIDE object describing that selection. Downstream nodes that accept an override then apply their behavior only to those layers. It's a plumbing node - the intelligence is in what consumes its output, not in the node itself.

The inputs and outputs that matter

There's exactly one input and one output:

• layers (STRING, multiline) - the list of attention layers to target, one specification per line. This is the entire interface. What values make sense depends on the technique you're feeding and the LTX checkpoint's block layout, so this is firmly an experimentation surface, not a set-and-forget default.

Output is a single ATTN_OVERRIDE that plugs into a compatible ltxtricks node.

How to install it

Install the whole pack via ComfyUI Manager (Ctrl+M → Install Custom Nodes → search LTXVideo → Install → restart), which is the README's recommended path. Or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/Lightricks/ComfyUI-LTXVideo

and restart ComfyUI. The node sits under ltxtricks → attn. It downloads nothing itself, but the trick nodes it feeds run on the LTX video model, which Lightricks specs at 32GB+ VRAM and 100GB+ disk.

Common issues & troubleshooting

It doesn't seem to do anything. Correct - by itself it doesn't. LTXAttnOverride only matters when its ATTN_OVERRIDE output is wired into a node that consumes one. If your graph doesn't route it somewhere, it's a no-op.

You have no idea what to put in layers. You're not alone; this is the least documented, most advanced part of the pack, and it assumes familiarity with the model's internal block structure. The realistic move is to start from a working ltxtricks example graph and vary the selection from there, rather than guessing layer indices cold. The original LTXTricks repo and the community around these editing tricks (the Banodoco Discord is where a lot of that experimentation happened) are your best reference.

Errors when it's parsed. The text has to match the format the node expects - malformed lines will throw. Keep the entries clean and consistent with whatever the example workflow used.

Honestly, if you're not already doing RF-inversion or attention-injection style editing on LTX, this node isn't a starting point - it's an advanced accessory. Get a basic LTX generation working first, then reach for the ltxtricks family when you specifically need to edit or steer existing footage.