🎲 LTX Tiled Sampler

This is the node that actually solves the LTX2 upscale problem, and it's the one I'd reach for first out of the whole pack. When you run a refinement pass on a 2× upscaled latent, you're feeding the DiT 4× the spatial tokens it was trained with. Every token now gets roughly a quarter of the per-token text-conditioning influence it expects, and the model starts operating outside its trained token-count range. The result is what you've seen a hundred times: broad hue shifts, color drift, and the prompt quietly losing its grip.

LTX Tiled Sampler is a drop-in replacement for SamplerCustomAdvanced that refuses to play that game. It splits the latent spatially along its longer axis with overlap, samples each tile as its own standalone clip at training-distribution token count, then blends the tiles back with a cosine-windowed Hann overlap. Same inputs, same wiring, better behavior.

How it works

Each tile runs the full sampling pipeline independently - noise, guider, sampler, sigmas, all of it - then the results blend. Don't confuse it with MultiDiffusion-style per-step coordination; this is whole-clip tiling, which is why it only works for light refinement. Every tile is a coherent sample on its own, and cosine blending hides the seams between them. It's transparent to your existing wiring: feed it the same noise source, guider, sampler object, and sigmas you'd give a normal sampler.

The inputs, beyond the standard sampler set (noise, guider, sampler, sigmas, latent_image):

• tile_axis - auto splits the longer dimension (right call for most portrait/landscape video); override with H or W if you have a reason.
• n_tiles (default 2) - tiles along the chosen axis. 2 suits most aspect ratios; go 3–4 for 4K+ outputs. 1 effectively disables tiling.
• tile_overlap (default 8) - overlap in latent tokens. 8 reliably hides seams; shave it only if VRAM is tight, and expect seams if you do.
• max_size_for_no_tile (default 24) - auto-skip tiling when the axis is at or below 24 latent tokens (≈768 pixels at LTX's 32× VAE compression). Inputs in the model's comfort zone just sample normally.
• bypass_tiling - the honest escape hatch: routes to standard single-pass sampling so you can A/B against a non-tiled baseline or debug whether a problem is tile-related at all.
• audio_pass - LTX2 is a video+audio model, and here's the clever bit. passthrough (default) leaves audio untouched. tile_carrying runs audio sampling alongside one chosen tile - the audio_carrier_tile (first, middle, last) - so video-audio cross-attention (read: lipsync) happens naturally during that tile's pass, at zero extra sampling steps. For talking-face content, first is the one: the speaker's usually in the top half.

Outputs are output and denoised_output, both LATENT, matching SamplerCustomAdvanced's contract.

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.

The one rule

Light denoise, few steps, refinement only. This is not a first-pass-from-noise sampler. The README is blunt: heavy-denoise tiling produces tile divergence that cosine blending can't reconcile. Typical upscale-pass schedule: sigmas starting moderate (around 0.85) with ~3 active steps, low denoise. The tooltip also flags euler_ancestral_cfg_pp as the recommended sampler algorithm for LTX2's distilled CFG=1 setups. Follow that and you get the color-stable upscale pass that base LTX2 refuses to give you for free.