LTX Attn — QKV Multiplier

A ComfyUI node in g_raw/LTX/Profiler with 8 inputs and 1 output.

By g-raw·Created 2 months ago·Updated 29 days ago· 2

model

steered_model

◄targets24:8►

◄apply_satrue►

◄apply_catrue►

◄qk_mult1.00►

◄vo_mult1.00►

◄from_step0►

◄to_step999►

Categoryg_raw/LTX/Profiler

Inputs (8)

Name	Type	Default	Description
model	MODEL		—
targets	STRING	24:8	Same format as Head Freeze/QKV Transfer's targets: paste Head Candidates' candidates_csv directly (one 'block,head' per line), or type manually as 'block:head \| block:h1,h2,... \| block:all \| ...'. A whole-string 'all' (or 'all:all') targets every block (0-47) and head (0-31). No prior capture needed -- this is a live multiply, not a replay. Leave blank to disable entirely -- this is the reliable way to turn it off; ComfyUI's node bypass/mute skips this node's cleanup, so the diffusion_model (shared across runs) can be left patched from a previous run.
apply_sa	BOOLEAN	true	—
apply_ca	BOOLEAN	true	—
qk_mult	FLOAT	1.00-10–10	Scales Q for the targeted heads before the attention dot product -- changes attention sharpness (softmax logit scale), not the head's output magnitude. 0 makes the head attend uniformly over all keys, it does NOT ablate it (still contributes via V). Scaling Q and K separately would be redundant -- both are uniform per-head scalars, so only their product changes the logits, hence a single knob.
vo_mult	FLOAT	1.00-10–10	Scales the targeted head's output (V before the attn_weights@V matmul, equivalently the head's slice of the output after -- same net effect either way since the matmul is linear in V, hence a single knob). Directly scales the head's contribution to the residual stream. 0 zeroes it out (true ablation).
from_step	INT	00–999	Denoising step (per targeted block) to start applying from. Defaults to the full range.
to_step	INT	9990–999	Denoising step (per targeted block) to stop applying at, inclusive. Defaults to the full range.

Outputs (1)

Name	Type	Description
steered_model	MODEL	—