DaSiWa LTX-2 LoRA Loader

Most LoRA stackers assume a model with one thing to condition: pixels. LTX-2 (and its current release, LTX-2.3) doesn't work that way - it's Lightricks' audio-video model, generating both from separate transformer branches inside a single 19–22B model. A LoRA trained to change how people move can end up also mangling the audio branch if you apply it at a flat strength across both, and that's the exact problem this node exists to solve: independent strength control per branch, per LoRA slot.

How it works

You get 10 stacking slots. Each one holds a LoRA file, a master STR multiplier (−2.0 to +2.0), and two branch-specific multipliers: V× for the video transformer, A× for the audio transformer. Scan a LoRA and the node shows you its video/audio key counts before you even queue a run, so you can tell at a glance whether a given LoRA is going to touch both branches or just one - a LoRA with zero audio keys doesn't need its A× multiplier touched at all. The whole thing lives on one node with a searchable LoRA list and six color themes (Jade, Neon, Studio, Chrome, OLED, Wood) if you care about matching it to the rest of your graph.

The provenance matters here: the loader is built on top of Brojakhoeman's Loradaddyloaderltx rather than invented from scratch, so the stacking mechanics are the same ones the LTX community was already using - DaSiWa adds the dual-branch split and the UI polish on top.

The inputs and outputs that matter

The node's formal inputs are simple - the stacking UI does the heavy lifting:

• model (MODEL) and clip (CLIP) - your loaded LTX-2/2.3 checkpoint, wired in the usual place before your sampler.
• stack_data - the serialized state of everything you configure in the node's widget (which LoRAs are loaded, their strengths, V×/A× per slot). You don't hand-edit this; it's what the UI writes as you add LoRAs through the node itself.

Outputs are the patched MODEL and CLIP, which you wire straight into your sampler and text encoder exactly like a standard LoraLoader - this node is a drop-in replacement for that step, not an extra stage.

How to install it

ComfyUI Manager - search DaSiWa-Nodes, install, restart.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/darksidewalker/ComfyUI-DaSiWa-Nodes
pip install -r ComfyUI-DaSiWa-Nodes/requirements.txt

No LTX model files ship with the node - you still need LTX-2/2.3 checkpoints and your LoRAs in the normal models/checkpoints and models/loras folders; this node just changes how you apply the LoRAs once they're there.

Common issues & troubleshooting

A LoRA does nothing. Check the key-count indicator first - if it shows zero keys for the branch you're trying to affect, the LoRA simply wasn't trained on that branch, and no amount of turning up V× or A× will change that. This is also the first thing to check before assuming the LoRA itself is broken.

Video looks right but audio is wrecked, or vice versa. That's exactly the failure mode this node was built to prevent - you're almost certainly running an older flat-strength loader instead, or you've cranked one branch's multiplier without checking the other. Bring the offending branch's multiplier down independently rather than reducing STR across both.

Distilled LTX-2.3 checkpoints are picky about LoRA strength. This isn't specific to this loader - it's an LTX-2.3 characteristic: the community-settled range for the distilled model in two-stage workflows is around 0.4 on the first sampler pass and 0.5 on the second, which is a useful starting point before you touch this node's own multipliers.