🅛🅣🅧 LTXV Tiled Sampler

LTXV Tiled Sampler is how you push LTX to a resolution your VRAM can't hold in one shot. Instead of denoising the whole frame at once, it slices each frame into a grid of overlapping tiles, samples them, and blends the seams back together. If you've ever hit an out-of-memory wall trying to render at 1080p or above, this is the node that trades a bit of time for a much bigger canvas.

It's part of ComfyUI-LTXVideo, Lightricks' official pack. Spatial tiling is an old ComfyUI idea, but doing it on video - where every tile also has to stay temporally consistent - is fiddly, and this node bakes in the blending so you don't have to hand-build it. It's the kind of workaround the LTX community leans on hard: people routinely report squeezing high-res LTX-2.3 out of 16GB cards by splitting the work like this.

How it works

You set a horizontal_tiles × vertical_tiles grid; the sampler runs LTX on each tile with a small overlap border, then feathers the overlaps together so you don't see grid lines. Peak VRAM is set by the size of one tile, not the whole frame, which is what lets a modest card render a big picture. The trade is real, though - more tiles means more passes, so more time, and aggressive tiling can soften fine detail or introduce faint seams if the overlap is too thin.

The inputs that matter

Most of the required sockets are the standard LTX sampling stack - model, vae, noise, sampler, sigmas, guider, and the input latents to sample into. The tiling controls are the reason you're here:

• horizontal_tiles / vertical_tiles - the grid, each 1–6. Start at 2×2 and only go higher if you're still OOM.
• overlap - the shared border between tiles (default 1). More overlap blends cleaner but costs a little speed; too little and seams can show.
• latents_cond_strength - default 0.15, how much the tiles lean on the shared latent for cross-tile consistency.
• boost_latent_similarity - a toggle that pushes tiles to agree more, trading a touch of variety for fewer seams.

For image conditioning there's optional_cond_images, optional_cond_indices, and images_cond_strengths, plus a crop mode. Two outputs come back: output and denoised_output, both LATENT - feed one into VAE Decode.

Installing it

ComfyUI Manager → Install Custom Nodes → search LTXVideo → install → restart. Or manually: cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo then restart. It appears under sampling. As with the rest of the pack, the LTX checkpoint and (on LTX-2) the Gemma text encoder are the big downloads, pulled on first use - budget the 100GB+ disk the pack asks for.

Where people get burned

• Reaching for tiles you don't need. If your resolution already fits in VRAM, a plain LTXV Base Sampler is faster and cleaner. Tiling is a memory workaround, not a quality upgrade - it can slightly reduce fine detail.
• Overlap too small. Thin overlaps at high tile counts are where visible seams and color mismatches creep in. Bump overlap before assuming the model is broken.
• Confusing spatial and temporal. This node tiles within a frame to reach higher resolution. If your problem is length - you want a longer clip, not a bigger one - that's LTXV Looping Sampler's job, which tiles across time instead.
• System RAM. On LTX-2/2.3, the gating resource for low-VRAM setups is usually system RAM, not VRAM. Nearly every "I ran it on 16GB" report pairs with 64GB of system memory; if you're swapping to disk, that's your slowdown.