LTX Flow Edit CFG Guider

This is the node behind prompt-based video editing on LTX - change "a man walking in a city" into "a man walking on the moon" while keeping the same motion, framing, and structure. It does it with FlowEdit, an inversion-free editing method, which is the important part: you don't have to invert the source video back to noise first, which is the slow, fragile step that older editing approaches all depended on.

How it works

Normal editing pipelines invert your source into a latent, then re-generate toward a new prompt, and inversion is where quality leaks out. FlowEdit skips it. It computes a direct path in the model's flow space from what the video currently is to what you want it to be, guided by two prompt pairs at once. You give it a source description and a target description, and the guider blends them so the parts you didn't mention stay put while the parts you changed move. FlowEdit first showed up in the image world with Flux (via Fluxtapoz) and got ported to video; this is the LTX flavor. The output is a GUIDER for a SamplerCustomAdvanced.

The inputs that matter

• source_pos / source_neg - conditioning describing the video as it is. Be accurate; this is the anchor.
• target_pos / target_neg - conditioning for what you want it to become. Change only what you want changed.
• source_cfg (default 2) - guidance on the source side.
• target_cfg (default 4.5) - guidance on the target side. The gap between the two is roughly "how much editing pressure" you apply; a bigger target_cfg pushes harder toward the new prompt.

It also takes the model. That's the whole surface - it's a focused node.

Installing it

• ComfyUI Manager - search LTXVideo, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo, then restart.

You'll also want a forward-prediction model patch in the graph (LTX ships one) since flow-edit runs the model in that direction - the example editing workflows wire it up for you.

Common issues

The edit changed too much - the whole scene drifted. Your source prompt probably didn't describe the video well enough, so FlowEdit had nothing to hold fixed. Tighten source_pos, and lower target_cfg to ease off the editing pressure.

The edit barely took. The opposite lever: raise target_cfg (or widen the gap between source and target). If it's still stubborn, your two prompts may be too similar for the guider to find a direction.

Motion breaks up on a big change. FlowEdit preserves structure best on edits that don't fight the original motion. A dramatic subject swap is harder than a style or setting change - and on LTX specifically, fast motion already smudges, so keep the edit within what the source clip can support.