Nodes/ComfyUI-LTXVideo/πŸ…›πŸ…£πŸ…§ Linear transition with overlap
ComfyUI Node Runs on cloud

πŸ…›πŸ…£πŸ…§ Linear transition with overlap

Crossfade video latents to stitch clips without a seam

By LightricksΒ·Created 2 years agoΒ·Updated about a month agoΒ· 3,956
πŸ…›πŸ…£πŸ…§ Linear transition with overlap
  • samples1
  • samples2
  • LATENT
β—„overlap1β–Ί
β—„axis0β–Ί

When you build a longer video out of several short LTX generations, the hard part isn't making the clips - LTX cranks those out in seconds - it's joining them without a visible jolt at each cut. LinearOverlapLatentTransition solves that by crossfading two latent clips over an overlapping window rather than butting them together. The seam turns into a short dissolve, and the eye stops catching it.

The key detail is that it happens in latent space, before decoding. You're blending the model's compressed representation, so the transition stays coherent with how LTX encodes motion and detail - much better than trying to feather decoded frames in a video editor afterward. If your workflow is "generate segment, generate next segment, chain them," this is the glue.

How it works

Give it two latent sequences and it interpolates across the overlap: the outgoing clip's tail fades down while the incoming clip's head fades up, linearly, across the frames you specify. The two are then merged into one continuous latent. Linear means no fancy easing - a straight, predictable ramp from one clip to the other, which is exactly what you want for an invisible join.

The inputs and outputs that matter

  • samples1 / samples2 (LATENT) - the two clips in play order. First is outgoing, second incoming.
  • overlap (INT, default 1, range 1–256) - the length of the crossfade in latent frames, and the only setting you'll really tune. Small overlap β‰ˆ a near-instant cut; larger overlap β‰ˆ a longer, gentler dissolve at the cost of frames from both ends. These are compressed latent frames, so even a handful covers a noticeable stretch of real video.
  • axis (INT, optional, default 0) - the dimension to blend along; 0 is the time axis, which is what a temporal transition needs. Leave it be unless you're doing something unusual.

Out comes one blended LATENT, ready for the VAE decode or another transition to add a third segment.

How to install it

Per the README, the quick way is ComfyUI Manager: press Ctrl+M, choose Install Custom Nodes, search LTXVideo, install, and restart. By hand:

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

then restart. You'll find it under LTXVideo β†’ latent. The node is trivial to run; the weight is in the LTX-2 model it serves, which Lightricks lists at 32GB+ VRAM and 100GB+ disk. First run downloads the required models.

Common issues & troubleshooting

Shapes don't match. The two latents must agree on every dimension except the one being blended. Different resolution or frame layout between clips will error out - keep generation settings identical across segments.

A see-through, double-exposed middle. Too long an overlap between clips that don't share much. A linear blend shows both at once during the dissolve, so wildly different scenes ghost. Trim overlap, or generate the incoming clip to begin close to where the previous one ended for smoother continuity.

No visible smoothing at all. With overlap at 1 there's almost nothing to fade. Increase it. Also double-check axis is 0 so you're blending time, not space.

Note: this pack ships a near-identical LTXVLinearOverlapLatentTransition (same inputs, same behavior, LTXV prefix). If both appear in your node search, pick either - they do the same thing.

CategoryLightricks/latent

Inputs (4)

NameTypeDefaultDescription
samples1LATENTβ€”
samples2LATENTβ€”
overlapINT11–256β€”
axisoptINT0β€”

Outputs (1)

NameTypeDescription
LATENTLATENTβ€”