(Deno) LTX High resolution Tiled Sampler

This is the node for the moment LTX 2.3 gets good but too big. LTX 2.3 is a 22B model, and its whole selling point is synchronized video and audio in one latent. That's great until you want a second refinement pass at high resolution, at which point a single full-resolution sample step can demand more VRAM than you have. The (Deno) LTX High resolution Tiled Sampler splits the video half of the latent into overlapping spatial tiles, evaluates the model per tile, fuses the predictions back together, and only then lets the sampler update - one global trajectory, no tile-drift chaos.

How it works

The clever bit is that it's not just "tiled sampling." The node keeps a single global sampler trajectory while video predictions run through the tiles, and the full audio latent is passed to every video tile as context. That way the video gets denoised against the audio for sync, but the audio itself is never re-denoised - in audio_mode freeze (the only mode and the default) the sound comes back untouched. If you're doing a low-denoise refinement pass to add resolution and detail, you want your existing audio preserved exactly, so this is the right trade.

Guide-bearing LTX latents are handled too: the tiling keeps guide metadata intact, but the node is explicit that for those you should run LTXVSeparateAVLatent first, then apply LTXVCropGuides to the video latent before decode.

The inputs that matter

• noise, guider, sampler, sigmas - these come from the standard ComfyUI sampler stack. The guider is BasicGuider or CFGGuider, and it should carry the LTX guide metadata from your upscaler/refinement path. sigmas is the low-denoise schedule for the second pass.
• latent_image - the LTX AV nested latent containing video and audio.
• horizontal_tiles / vertical_tiles - split counts. 2×2 is a sane start; push higher only if you're still OOMing.
• overlap - overlap in latent tokens between tiles. Default 8; more overlap means smoother blending but more compute.
• blend_mode - hann (default and recommended) is where you start.
• aggressive_memory_cleanup - on by default. It runs extra cleanup between tile predictions; the tooltip is honest that this can be much slower, but it helps fragmented-memory cards.

Outputs are output (the refined AV latent, audio preserved) and denoised_output (the x0 callback latent), so you can pick the path your workflow needs.

Install and context

The node ships in the Deno pack - Manager search "Deno Custom Nodes" or:

cd ComfyUI/custom_nodes
git clone https://github.com/Deno2026/comfyui-deno-custom-nodes.git
# restart ComfyUI

It's a sampler that plugs into the official ComfyUI-LTXVideo workflow, so you need those LTX nodes (and their 2.3 models) installed as the base. If you don't have the LTX 2.3 setup already, the (Deno) LTX Model Loader and (Deno) Easy Model Download Helper in the same pack exist to get you there - the 2.3 GGUF starter set is the built-in preset for lower-VRAM rigs.

Two things trip people up. First, this node is for AV latents - if your latent is video-only, the (Deno) LTX Tiled Spatial Upscaler is the right sibling, and feeding a plain video latent here will complain. Second, don't expect speed; tiling is the price you pay to fit, and aggressive_memory_cleanup makes it slower still. Leave it on if you're on a 12–16GB card, and remember LTX 2.3's rebuilt VAE means previews are unreliable anyway - you're often working blind until decode, which is exactly when a node that fits without crashing earns its keep.