GU Set Frames Video

GU Set Frames Video exists to enforce one boring, essential formula: video frame counts that must satisfy N*8 + 1. LTX and WAN-style pipelines want frame counts in that shape - 9, 17, 25, 33, 121, 257... - because of how the temporal layers chunk the video. This node gives you a slider that can only produce those numbers, so you never feed the model a frame count it has to reject or, worse, silently pad into something you didn't intend.

It's the video analog of GU Resolution Select: a guardrail disguised as a convenience.

What it does

One input, frames, an integer from 9 to 1001 stepping by 8 - so every stop on the slider is automatically of the form N*8 + 1. The tooltip literally reads N*8 + 1. One output, an INT, which you wire into the frame-count input of your video latent or video pipeline.

Mechanically there is nothing else: the node validates the shape on the frontend and passes the number straight through. The whole job is preventing the invalid counts in the first place.

Why it matters in practice

Different video models have different frame-shape rules, and getting them wrong is a real failure mode. The LTX line is the clearest example - the KB's LTX entity records the 2B original shipping 121 frames as its canonical clip and notes practice settling on ~121-frame clips, with LTX's architecture stepping in the same +1 pattern. WAN's native context is 81 frames (also 10*8 + 1), and the 1050 Ti run that got everyone talking was 97 frames (12*8 + 1). When you're chaining LTX drafts into WAN refinement - the standard "draft fast, refine slow" workflow the KB documents - keeping both models' frame counts in the right shape is exactly the kind of thing that bites you at the end of a long pipeline. This node means the number you hand the model is always legal, and the slider makes the relationship to clip length visible: each step is 8 more frames.

The README frames it as "LTX / WAN-style pipelines," and that's the honest scope - it's not a generic frame tool, it's the one formula.

Install

ComfyUI Manager → Custom Nodes Manager → search "GU Nodepack" → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/alexguryev/ComfyUI-GU_Nodepack
cd ComfyUI-GU_Nodepack
pip install -r requirements.txt

Deps: psutil, requests, gu-funclib>=1.9.0. No model downloads. Windows-tested, single-user design.

The honest notes

This is a thin node - if you're comfortable doing frames*8 + 1 in your head or with any math node, you don't strictly need it. Its value is that it makes the constraint visible and unbreakable on a slider, which is worth more than it sounds when you're mid-workflow at 1 a.m. One caveat: frame count isn't the only constraint the newer models impose. LTX-2 defaults to 24fps (48fps suggested for I2V), WAN wants certain resolutions with its counts - so treat this as the frame-count half of a checklist, not the whole thing. And for still-image sizing, pair it with GU Resolution Select.