✨ S42 CutFlow LTX Post-Process

If you've run LTX 2.3 with guide-frame conditioning (the LTXVAddGuide first/last-frame workflow), you know exactly what this node is for: the last frame or two of your generated video are garbage. The guide frame leaks through - a frozen, artifact-y end frame that looks nothing like the rest of the clip. Native ComfyUI has LTXVCropGuides to fix this, and S42CF_LTXPostProcess does the same job plus two bonus cleanup passes, all in one node you can drop between your VAE decode and your editing chain.

It's from S42 CutFlow's LTX Bridge group, GeekyGhost's set of nodes for plugging his editing suite into the LTX 2.3 pipeline. If you generate with LTX and then edit with CutFlow, this is the handshake node - it makes the raw generation safe to feed into everything else.

The inputs that matter

• video (IMAGE) - your LTX VAE decode output (the tooltip notes it'll also take an RTX-upscaled result if that's how you roll).
• crop_last_frames (int, default 2) - drops N frames off the end. Default 2 deliberately matches what LTXVCropGuides does, because that's where guide artifacts land. If you see a third bad frame, bump it to 3.
• crop_first_frames (int, default 0) - usually leave this alone. Only raise it if you used some unusual conditioning that poisons the start.
• auto_brightness - none (default), subtle, or moderate. Subtle nudges an over- or under-exposed clip back toward sane territory; moderate does a per-frame min/max auto-levels stretch. If your LTX output already looks good, leave it at none - the auto-levels stretch in particular can flatten a clip that was fine.
• denoise_strength (float, 0–1, default 0) - light temporal denoising. 0.3 is the "subtle" sweet spot the tooltip suggests. It's a neighbor-frame average, so it trades a little sharpness for smoothing out LTX's frame-to-frame sparkle.

How it actually works

The mechanism is refreshingly simple and fully deterministic. First it slices the frame batch - start = crop_first_frames, end = n - crop_last_frames - so the artifacts physically leave the tensor. Then the optional brightness pass, then a three-frame temporal blend when denoise_strength > 0: each middle frame gets mixed with the average of its neighbors. Nothing here loads a model, so it runs at near-zero cost and you can put it in a long pipeline without worrying about GPU pressure.

Outputs are video (the cleaned IMAGE batch), frame_count (INT - useful to verify your final length after cropping), and info (STRING), which tells you exactly how many frames went in and came out.

Installing it

Same as every other S42 CutFlow node: ComfyUI Manager, search S42 CutFlow, install, restart. Or clone it manually:

cd ComfyUI/custom_nodes/
git clone https://github.com/GeekyGhost/S42-CutFlow.git
pip install -r S42-CutFlow/requirements.txt   # Windows portable: use python_embeded\python.exe -m pip

Then restart. If the node shows up under S42 CutFlow/LTX Bridge in the menu, you're good.

Gotchas

The main trap is leaving crop_last_frames at 0 because you didn't realize the artifacts were guide leakage rather than a bad generation - if your LTX output consistently has one or two weird frozen frames at the end, this is the fix, not a re-roll with a higher step count. If you also run a native LTXVCropGuides node in your workflow, don't stack both with defaults, or you'll eat four frames instead of two - pick one. And remember the ordering: crop first, then grade. Feeding artifact frames into a Style Transfer or a sharpen pass just bakes the junk in.