🔍 LTX Latent Upsampler (Tiled)

LTX Latent Upsampler (Tiled) is the "wait, it was never the upscaler's fault" node. It's a drop-in replacement for ComfyUI's stock LTXVLatentUpsampler that processes the upscale in spatial tiles instead of all at once. The original reason was fixing color shifts and distortion at extreme aspect ratios - and then the pack author's own investigation found the upscaler wasn't actually the source of upscale-pass color drift. The sampler operating on upscaled token counts was (that's what the LTX Tiled Sampler fixes).

So what's left? A genuinely useful memory-safety tool. For big or extreme-aspect latents, tiling the upscale keeps only one tile in VRAM at a time, so the same upscale that OOMs on a portrait 1080×1920 latent just works. In safe-aspect cases the output is virtually identical to the non-tiled version, which is exactly what you want from a drop-in.

How it works

The upscale model runs on tiles with cosine-windowed overlap blending, and the node handles the bookkeeping for you. It un-normalizes the latent once using the VAE's per-channel statistics, tiles and upscales, then re-normalizes once on the full output - so you don't get the per-tile normalization seams that cheaper tiling implementations leak.

The hidden gem is rotate_for_landscape. The theory, per the source: the upscale model's training distribution is biased toward landscape (most video data is wider than tall), so tall portrait inputs are out-of-distribution and drift. Enable it and the node transposes H/W before the upscale forward - making your portrait a "landscape" for the model - then transposes back. It's a lossless geometric transform, and for portrait work it's the difference between clean upscales and color drift. I'd leave it on for anything portrait.

The inputs that matter

• samples, upscale_model, vae - required. The upscale_model is the LATENT_UPSCALE_MODEL from an Upscale Model Loader, vae is your LTX2 VAE.
• tile_size (default 24) and overlap (default 8) - tile geometry in latent tokens. If overlap is >= tile_size the node clamps it for you.
• max_size_for_no_tile (default 32) - inputs at or below this skip tiling entirely and run the stock non-tiled path, so small latents don't pay the tiling tax.
• debug - verbose console output for first-time verification.

Output is a single LATENT, ready for your refinement sampler (the tiled one pairs naturally).

Installing it

Same TenStrip pack:

cd ComfyUI/custom_nodes
git clone https://github.com/TenStrip/10S-Comfy-nodes.git 10S_Nodes

or ComfyUI Manager → "10S". Restart, no extra dependencies.

When to actually use it

Reach for this when your latent is big or tall and the stock upsampler chokes or drifts: extreme portrait aspects, 4K+ targets, or 12–16GB VRAM cards where a full-latent upscale forward is asking for trouble. If you're on a 24GB card at normal aspects, the stock LTXVLatentUpsampler is fine and you can skip this - but once your upscale pass starts throwing OOM or your vertical videos start looking like a different color grade, remember it was never the upscaler's fault, and this node exists to keep the upscaler honest anyway.