LTX Attn — QKV Multiplier

Every other intervention node in this pack needs a prior capture - the whole two-run dance. QKV Multiplier doesn't. It's a live multiply: scale the attention sharpness and/or output magnitude of targeted heads while the generation runs, no store, no setup. That makes it the fastest way to test "what if I kill this head?" - the profiling-then-freeze loop collapses into a single run.

Its two knobs are also the pack's clearest lesson in what attention math actually does, so this node doubles as a crash course:

• qk_mult rescales the softmax logits before attention weights are computed. It changes how sharp or flat the head's attention distribution is. Setting it to 0 does not ablate the head - it makes it attend uniformly over every key, and it still contributes via V. You made it unfocused, not dead.
• vo_mult scales the head's actual contribution to the residual stream - V before the attn_weights @ V matmul, equivalently the head's slice of the output after (linear in V, so same net effect). Setting it to 0 is a true ablation: that head genuinely stops contributing.

So: vo_mult to kill or amplify a head; qk_mult to experiment with attention sharpness. The tooltips in the node spell this out, and it's worth trusting them over your first instinct. There's deliberately no separate q_mult/k_mult or v_mult/o_mult pair - Q and K are each a uniform per-head scalar, so only their product shows up in the logits, hence one qk_mult; same logic collapses V/O into vo_mult.

The inputs that matter

• targets - same format as Head Freeze/QKV Transfer: block:head | block:h1,h2 | block:all | all. A whole-string all targets every block (0–47) and head (0–31). Multipliers are shared across all targets.
• apply_sa / apply_ca - which attention types get scaled.
• qk_mult / vo_mult - the two knobs above, range −10 to 10, default 1.0 (no-op). Note negatives are allowed - you can invert a head's contribution, which is a very particular kind of torture test.
• from_step / to_step - the denoising step window the multiply is active for (defaults to full range 0–999).

Output is a single steered_model MODEL.

Disable by blanking

Same house rule as the other intervention nodes: clear targets to disable, don't use bypass/mute. This node patches the shared diffusion_model, and bypass/mute skips the code that unwraps its layer, leaving the multiplier in effect on later runs while looking disabled. Blank targets runs the cleanup reliably. LTX Reset Patches is the sweep if you've orphaned a layer.

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, no capture store required - this is the lowest-friction node in the pack to try. Work-in-progress caveat applies to the pack as a whole.

Common issues

The main trap is conceptual, not mechanical: reach for vo_mult=0 when you want to ablate a head, because qk_mult=0 just makes it a uniform attender and the result confuses people ("I zeroed it and nothing changed" - right, it's still contributing). Because it's live, it also stacks with other intervention nodes if you chain them - that's supported (each node registers its own layer), but remember each one also needs its own disable-by-blanking when you're done.