🎚️ LTX LoRA Stack (AV)

If you've ever stacked three LoRAs on an LTX2 workflow, you know the pain this node kills: a sideways chain of LoRA Loader Advanced nodes, each a separate step, and if one is out of order the whole thing's a mess. LTX LoRA Stack (AV) is one node that holds up to twelve LoRAs at once, and it throws in a trick the loaders can't do - separate strength per LoRA for the audio half of the model versus the video half.

That AV split is the reason this node exists. LTX2 is a dual-stream DiT: video tokens and audio tokens flow through the same transformer but hit different block groups (audio_attn, audio_ff, audio_adaln on one side, the main transformer_blocks on the other). A LoRA can touch both, or mostly one. The stack lets you say "this style LoRA hits video at 0.8 but I want zero audio impact" without training anything.

How it works

Each LoRA you load is split into two buckets by a simple rule: any model key containing the substring audio lands in the audio bucket, everything else in the video bucket. That rule catches more than you'd think - audio_embeddings_connector, audio_to_video_attn, and video_to_audio_attn all contain "audio", so they're correctly treated as audio-side layers even though one of them is cross-attention from video.

Per layer, the effective strength is:

effective = strength × (audio_weight  for audio layers
                       video_weight  for video layers)

Each slot gives you three knobs: strength (the overall multiplier, −10 to 10), audio_weight, and video_weight. Set a side to 0 and that LoRA's influence on that side disappears. The node clones your ModelPatcher before applying, so the upstream model isn't mutated - standard Comfy good manners.

The inputs that matter

• model - the MODEL from your checkpoint loader.
• num_loras - the single source of truth for how many slots are active. This is the one beginners get burned by: slots beyond num_loras are ignored entirely, no matter what's in them. Bump it as you add LoRAs.
• lora_1 … lora_12 - dropdowns populated from your models/loras folder. Set one to "None" (or its strength to 0) to disable that slot.
• strength_N / audio_weight_N / video_weight_N - the three per-slot multipliers above.

Output is a single MODEL, wired straight into your sampler. The description on the socket couldn't be plainer: you're replacing a chain of loaders.

Installing it

This is part of the 10S-Comfy-nodes pack (TenStrip). Easiest is ComfyUI Manager - search "10S" and install. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/TenStrip/10S-Comfy-nodes.git 10S_Nodes

Restart ComfyUI and the nodes appear under 10S Nodes/. There are no extra pip dependencies - the README is explicit that everything this pack needs already ships with ComfyUI. Update with git pull in that folder.

Where people get burned

The num_loras trap above is the big one - a workflow file can carry stale values in slots you think are off. Also remember this pack is LTX2-specific: the LoRA key mapping assumes the LTX-AV class structure, so don't expect it to stack SDXL LoRAs. And because LTX2 is prompt-sensitive (the community's standing advice is long, specific prompts), a big stack of style LoRAs will amplify each other fast - start with strength 0.6–0.8 per LoRA and work up rather than slamming everything to 1.0.