LTX Attn — Reset Patches

You're not going to want this node - until you absolutely need it. Every intervention node in this pack (Setup Capture, Head Freeze, QKV Transfer, QKV Multiplier) patches the underlying diffusion model, and that model object is shared across every model.clone() in your session. The pack is careful about cleaning up its patches on normal runs, but there's one case nothing catches: you delete or rewire an intervention node out of the graph. Nothing calls cleanup for a node that stops executing, so its last-registered layer stays on the shared model - still consuming RAM, still applying on future runs, with zero visible trace. That's the stale-patch state, and Reset Patches is the only way to definitively clear it.

Think of it as rm -rf for the model's layer registry. It's also the escape hatch if you ignored the "blank targets to disable, not bypass/mute" rule and now have a freeze that "shouldn't be there" quietly running every generation.

How it works

It calls reset_all_layers() on the model's diffusion_model: every registered intervention/capture layer is cleared and the pristine _forward is restored. Nothing else survives - this isn't a selective undo, it's a full sweep, which is exactly what you want when you suspect an orphan.

One implementation detail worth knowing: this node has no widgets, only a MODEL input, so ComfyUI would cache it and skip re-running it whenever nothing upstream changed. Its whole job is a side effect, so the author made IS_CHANGED return float("nan") - NaN never equals itself, so ComfyUI is forced to treat it as changed and re-run it every queue. That's deliberate, and it means you can drop it anywhere and trust it fires.

The single input is model; the single output is clean_model (a clone with the layers stripped). Wire it downstream of the model - the README suggests anywhere downstream - and run the queue once.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/g-raw/ComfyUI-LTX-Attention-Toolkit.git

Restart ComfyUI. No extra dependencies, no model files. Work-in-progress pack, same caveat as the rest.

Common issues

The realistic failure is using it too late - after you've already chased a phantom. The correct habit, from the pack's own docs: if you've been iterating on a graph with intervention nodes and suspect any stale layer, run Reset Patches once before blaming your prompt or the model. And if you have intervention nodes in the graph and want them on, don't put Reset Patches in that path - it clears everything, including the live nodes' layers, and they'll re-register on their next run anyway, but you'll get one clean (unpatched) generation out of it that will confuse you.