LTX Forward Model Pred

This is a plumbing node - you won't build a workflow around it, but the flow-edit workflows won't run without it. It patches the LTX model to predict in the forward direction, which is the mode the FlowEdit editing method needs. If you've dropped in the LTX Flow Edit CFG Guider and something's off, odds are this is the piece you're missing.

How it works

Diffusion sampling normally runs one direction - from noise toward a clean image. Editing methods like FlowEdit need the model set up to reason about the path from an existing video to an edited one, which means changing how the model's sampling prediction is configured. This node does exactly that one thing: it takes a MODEL and returns a MODEL patched into forward-prediction mode. No parameters, no knobs - it's a mode switch you insert into the graph. Think of it like flipping a toggle the rest of the edit pipeline assumes is already flipped.

Because it has no settings, there's nothing to tune here. The interesting decisions all live in the guider you pair it with; this just puts the model in the right state for that guider to operate. That's typical of the low-level flow-edit nodes - the intelligence is spread across a few small single-purpose pieces rather than one big configurable node, which is more flexible to wire but means a missing piece fails quietly rather than loudly.

The input and output

• model - the LTX model to patch.

Output is the patched MODEL, which you route onward to your flow-edit sampler and guider. That's the entire surface - one in, one out.

Installing it

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

Common issues

Your FlowEdit workflow produces garbage or ignores the source. A missing or misplaced forward-pred patch is a prime suspect. The flow-edit guider expects the model to be in this mode; without it, the edit math is operating on a model that's set up for plain generation. Make sure this node sits in the model path feeding the flow-edit guider.

You added it to a normal generation and it got weird. Right - this is edit-pipeline machinery, not a general enhancer. In a standard text-to-video or image-to-video graph it has no business being there and will only confuse the sampler. Only reach for it when you're doing FlowEdit-style editing.

You're not sure it's doing anything. It won't show a visible effect on its own; it's a precondition, not an enhancement. The right test is whether the whole edit pipeline behaves, not whether this one node changes a preview.