LTX Reverse Model Pred

Small node, narrow job, and if you're doing LTX video editing you probably need it without quite knowing why. LTXReverseModelSamplingPred flips the direction of the model's prediction so it can run the sampling process backwards - from a clean video toward noise - instead of the usual noise-to-video direction. That backwards run is the inversion step that FlowEdit and RF-inversion editing depend on.

Here's the intuition. Normal diffusion sampling walks noise downhill into an image or video. To edit an existing clip, you first need to figure out what noise trajectory would have produced it. That's inversion: running the flow in reverse to recover the path. But the model was trained to predict in the forward direction, so you patch it to reverse its prediction. This node is that patch, for LTX specifically.

How it works

It takes your MODEL and returns a MODEL - same as any model-patching node in ComfyUI. Under the hood it wraps the model's sampling prediction so the sampler steps the reverse trajectory. You typically apply it to a copy of the model that drives the inversion/capture branch of an editing graph, while the forward branch uses the unpatched model. The reversed pass is what fills the attention bank and establishes the source's latent path; the forward pass then edits along it.

The inputs and outputs that matter

There's nothing to tune. One input, one output:

• model (MODEL) in - your LTX model.
• MODEL out - the reversed-prediction version, which you route into the inversion/unsampling half of your FlowEdit workflow.

No sliders, no seed, no strength. It's a behavioral flip, not a parameter. That's deliberate - the "how much" of an edit is controlled elsewhere (skip steps on the sampler, inject steps on the attention nodes), not here.

How to install it

ComfyUI Manager: Ctrl+M, Install Custom Nodes, search LTXVideo, install, restart. From the terminal:

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

then restart ComfyUI. It's in Lightricks' official LTXVideo pack, under the ltxtricks category. The node adds no downloads of its own; the LTX model it patches is the heavy part (the pack targets the LTX-2 family - plan for the full checkpoint, the Gemma 3 text encoder, and serious VRAM).

Common issues & troubleshooting

You applied it to the model that does the actual generation and got garbage. That's the classic mistake. The reversed model is for the inversion branch only - the pass that reads the source clip and captures its trajectory. Your final sampling should run on the normal, unpatched model. Wire two model branches, not one.

You're not sure whether you even need it. If your workflow already inverts the source some other way, or you're doing a lighter edit that doesn't reconstruct the source path, you may not. This node belongs to the full inversion-based editing pipeline. If you're just generating fresh video, skip it entirely.

The edit doesn't resemble the source at all. Inversion quality is fussy across all these methods - one FlowEdit user flatly called the whole family "a lot of black magic." If the reconstructed path is off, no amount of downstream tuning saves the edit. Confirm the inversion branch is wired end to end (reversed model → sampler → capture) before blaming the edit settings, and lean on a working example workflow rather than assembling this from memory.