LTX2 Lo RA Select (QQ)

LTX-2 and LTX-2.3 aren't just video models - they generate video and synchronized audio in one pass, which means their attention blocks come in flavors: pure video, video-to-audio cross-attention, audio-to-video, pure audio, and "other." A plain LoRA loader treats all of those as one blob. LTX2LoRASelect is the node that knows better: it loads an LTX LoRA but lets you set the strength independently for each attention type. That's the whole reason it exists, and it's a capability you simply don't get from the vanilla LoRA loader.

Inputs: lora_name (dropdown from ComfyUI/models/loras), model (the diffusion model to apply to), and strength_model (the overall multiplier, −100 to 100). Then the per-block dials, each 0–1 with a 1.0 default:

• video - strength for video attention layers.
• video_to_audio - video→audio cross-attention.
• audio - pure audio attention.
• audio_to_video - audio→video cross-attention.
• other - layers that don't fall into the above.

Output is a single MODEL, patched with the LoRA at those per-block strengths. The code docstring notes the node "outputs model with attached metadata for daisy-chaining multiple LoRA loaders" - so you can string several LTX2LoRASelect nodes together, each contributing its own LoRA with its own per-block profile, and the last one hands the fully-stacked model to the sampler.

How to think about the per-block strengths

This is the dial that matters for LTX specifically: a LoRA trained to change the voice lives mostly in audio layers, a motion LoRA lives in video layers, and a style LoRA touches both plus the cross-attention bridges. Being able to set audio to 0.8 while keeping video at 1.0 (or vice versa) is how you get a LoRA to change only what you want it to change. If you're porting an LTX-2.3 LoRA and it's changing the soundtrack when you only wanted motion, this node is the surgical fix. It also plays into the broader LTX story - Lightricks' engine-style licensing and speed-first niche aside, LTX-2.3 is where the current open-audio action is, and LoRAs here are the community's main way to shape it.

Installing it

Part of siraxe/ComfyUI-WanVideoWrapper_QQ via ComfyUI Manager (search WanVideoWrapper_QQ/SA-Nodes-QQ), or:

cd ComfyUI/custom_nodes
git clone https://github.com/siraxe/ComfyUI-WanVideoWrapper_QQ.git

Restart ComfyUI. You'll need an LTX checkpoint in ComfyUI/models/checkpoints and your LoRAs in ComfyUI/models/loras.

Gotchas

• Block strength ≠ overall strength. strength_model multiplies everything; the per-block sliders shape the ratio. If a LoRA seems inert, check both.
• Order matters when chaining - each node applies to the model it receives, so chain order = application order.
• Pack note: repo renamed to ComfyUI-SA-Nodes-QQ in v1.3.4; delete any stale wanwrapper_qq folder in custom_nodes if old workflows reference it.