LTX Director Motion Brush V2 Safe Downscale Factor
LTX Director Motion Brush V2 Safe Downscale Factor
- latent
- latent_downscale_factor
- summary
This is a tiny node with an even tinier job, and it deserves a sentence of respect before you dismiss it: it prevents a specific, maddening error. The IC-LoRA motion-track guide in ComfyUI-LTXVideo takes a latent_downscale_factor and dilates it against your latent grid. When the latent's width and height aren't divisible by that factor, the dilation throws - and it's not an obvious error, it's the kind that makes you rewire half a workflow before you realize the fix is "don't request an incompatible number."
So the node does exactly one sanity check: it looks at your latent, and only passes the factor through when the latent dimensions actually divide evenly. Otherwise it falls back to a safe value - default 1.0 - so the pipeline just works.
The three inputs
latent- the video latent you're about to sample.requested_factor(default 1) - what you'd like the downscale factor to be.fallback_factor(default 1) - what to use when the requested factor doesn't divide the latent cleanly.
That's the whole thing. You set a requested value, it checks divisibility, and it either gives you what you asked for or the fallback.
Outputs
latent_downscale_factor (a FLOAT you wire into the IC-LoRA guide node's downscale-factor input) and summary - a JSON string worth actually reading once. It reports the effective factor and, crucially, a reason field: requested_factor_compatible, latent_grid_not_divisible, fallback_factor_compatible, and so on. When your video comes out wrong-shaped or the guide node starts complaining, that summary tells you in one glance whether the factor silently became 1.0 and why.
Install and wiring
Part of the pack, no separate install:
comfy node install ltx-director-motion-brush
# or clone into ComfyUI/custom_nodes, then restart ComfyUI
Wire the latent from your video-latent source into latent, set requested_factor to what your IC-LoRA setup expects, and feed the FLOAT output into the guide's latent_downscale_factor input.
The honest take
You will never think about this node again once it's wired - and that's the point. It's the kind of guardrail a workflow ships with so you don't have to learn what "dilation error on non-divisible latent grid" means at 1 a.m. The only time you'd touch it deliberately is if you're pushing resolutions: if you're generating odd dimensions and the guide keeps failing, check the summary before you blame the sampler. Nine times out of ten it's the grid, and this node is telling you exactly that.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| latent | LATENT | — | |
| requested_factor | FLOAT | 11–10 | — |
| fallback_factor | FLOAT | 11–10 | — |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| latent_downscale_factor | FLOAT | — |
| summary | STRING | — |