🅛🅣🅧 LTXV Dilate Latent

A utility node, and a genuinely simple one: it takes an LTX latent and dilates it - spreads it out - by an integer grid factor in width and height. No pixels, no VAE, just the latent tensor getting expanded on a grid. It's the kind of node you never search for on purpose but end up needing inside a bigger LTX workflow, usually one doing spatial tricks or editing.

Why would you want to dilate a latent? A few reasons show up in practice. Video latents in LTX are spatially compressed, and some editing or reference tricks want the latent laid out on a coarser or finer grid before another operation runs on it. Dilating inserts spacing on that grid, which changes how a downstream node samples or aligns against it. It's a building block, not a result - think of it like a "resize on the grid" for latents rather than a creative control.

How it works

The node reads your LATENT, applies the horizontal and vertical scale factors to its spatial grid, and hands back a dilated LATENT. Because it's operating in latent space with integer grid factors, the effect is structural - it rearranges/expands the grid rather than smoothly interpolating an image. That's exactly why it lives in a workflow next to other latent-manipulation nodes and not on its own.

The inputs and outputs that matter

Three inputs, two of them the actual controls:

• horizontal_scale (INT, default 1, range 1–100) - the grid factor across width. At 1 (the default) nothing changes; 2 spreads the grid twice as wide, and so on.
• vertical_scale (INT, default 1, range 1–100) - same thing for height.
• latent (LATENT) - the tensor you're dilating.

Output: a single LATENT. Both scale defaults are 1, which means out of the box this node is a pass-through - you have to raise a factor for it to do anything.

How to install it

Via ComfyUI Manager: press Ctrl+M, choose Install Custom Nodes, search LTXVideo, install, restart. Or clone it:

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

and restart ComfyUI. Ships in Lightricks' official pack. This node is pure tensor math - it downloads nothing and costs almost no compute - but it's meant to sit inside an LTX graph whose model and text encoder are the heavy part.

Common issues & troubleshooting

Nothing happened. Both scales default to 1, i.e. identity. If you dropped the node in and left the defaults, it's a no-op by design. Set a factor above 1.

The output looks blocky or grid-y. That's inherent to grid dilation in latent space - it's a structural expansion, not a smooth upscale. If you wanted a clean resolution bump, this is the wrong tool; use LTX's actual spatial upscalers (the latent_upscale_models the pack ships) or a pixel-space resize instead.

A downstream node rejects the dilated latent. Big scale factors change the latent's shape, and not every node that follows expects the new dimensions. If you're getting shape-mismatch errors, back the factor down or check that whatever consumes the dilated latent is built to handle the expanded grid. This is a low-level building block - it trusts you to know why you're stretching the grid, so if you're not sure, you probably don't need it in the graph.