LTX Flow Edit Sampler

This is the node you reach for when you already have a video and you want to change it - repaint the subject, swap the setting, restyle the whole thing - while keeping the original motion and composition intact. It's LTX's take on FlowEdit, the "edit without inversion" trick that got popular on Flux and Hunyuan before landing here. If you've ever seen those "remix this clip at varying strengths" LTXV demos, this is the machinery under them.

The reason FlowEdit exists is that the obvious way to edit a generation - invert it back to noise, then re-sample with a new prompt - is slow and drifts. FlowEdit skips the round trip. It computes the difference between where the source prompt wants to go and where your target prompt wants to go, and pushes the latent along that difference in a single pass. As one person who's used both put it, FlowEdit "does the unsampling/sampling on the same pass so it's faster," and it tends to stay truer to the prompt and keep more consistency than RF-inversion. That's the trade you're buying into: faithful to the original, fast, a little less wild.

How it works

You don't wire a source video straight into this node. LTXFlowEditSampler outputs a SAMPLER - a sampling strategy object - which you feed into a custom-sampling graph (a SamplerCustom / SamplerCustomAdvanced setup) alongside your source and target conditioning and the model. The sampler is what knows to run the FlowEdit delta instead of a plain denoise. Everything about what you're editing lives in the surrounding graph; this node decides how the edit is applied.

The inputs and outputs that matter

Three inputs, and only one you'll really touch:

• skip_steps (default 4) - the fidelity dial. It controls how many steps of the edit trajectory get skipped, which in practice trades "stays close to the source" against "commits harder to the new prompt." Higher skip means a gentler, more faithful edit; lower means a bigger departure. This is the knob you tune first and tune most.
• refine_steps (default 0) - optional extra steps to clean up the result after the edit. Leave it at 0 to start; bump it if the output looks slightly undercooked.
• seed - the usual. Changing it re-rolls the noise on the edit.

Output is a single SAMPLER, straight into your custom sampler node.

How to install it

Through ComfyUI Manager: open Manager (Ctrl+M), Install Custom Nodes, search LTXVideo, install, restart. Or from the terminal:

cd ComfyUI/custom_nodes
git clone https://github.com/Lightricks/ComfyUI-LTXVideo

then restart ComfyUI. The pack is maintained by Lightricks themselves. Note the pack as a whole targets the LTX-2 family and expects the LTX checkpoints, the Gemma 3 text encoder, and 32GB+ VRAM with 100GB+ disk for models - but this particular node just patches sampling and downloads nothing on its own.

Common issues & troubleshooting

Your edit barely changed anything. Lower skip_steps. At the default of 4 the sampler is deliberately conservative to preserve the source; if you want a real transformation you often need to skip fewer steps.

Your edit went off the rails and lost the original motion. The opposite problem - raise skip_steps to keep more of the source trajectory.

You wired the video into this node and it errored. It doesn't take a video. It takes nothing but the three settings and emits a SAMPLER; the source and target conditioning belong on the custom-sampler node downstream. If your graph looks nothing like a standard SamplerCustom flow, start from a FlowEdit example workflow rather than building it from scratch - this is a "tricks" node, and the wiring is the hard part, not the settings.

Nothing's faithful to the prompt. Remember LTX is prompt-sensitive to begin with. FlowEdit inherits that: a short target prompt gives a mushy edit. Write the target the same way you'd write a fresh LTX generation - long and specific.