Gear Tiled Fusion Sampler (LTX)

LTX is the fast one. While Wan is still chewing on a 5-second clip, LTX hands you a result in seconds - which is exactly why its IC-LoRA video-to-video scene is so active. The catch is the training window. LTX models are trained at a modest native resolution, so a 4K upscale/refine pass either OOMs you or falls apart.

Gear Tiled Fusion Sampler (LTX) is the node that says "fine, sample the whole canvas anyway." It keeps one latent canvas and one global noise field, runs the model once per overlapping spatial tile at every denoise step, and Gaussian-blends the stepped tiles back onto the canvas. Because every tile shares the same canvas and noise, they ride one coupled trajectory - no visible tile boundaries, which is what normally kills naive tiled approaches.

Why you'd actually reach for it

This node exists for IC-LoRA video-to-video at resolutions far above what the model was trained on: upscaling/refining, SDR-to-HDR, day-to-night style transfers. Those are the jobs the Gear pack's HDR sibling nodes (the LogC/ACEScct decode-and-save-EXR family) feed and consume.

The subtle part is the guide. An IC-LoRA's conditioning is spatial - attention entries and keyframe indices attach to pixel positions on the full canvas. Tiling the latent isn't enough; each tile needs the crop of the guide that covers it. This node does that conditioning surgery per tile, then crops the appended guide frames off the output itself - so you don't need LTXVCropGuides anymore.

Wiring it up

It's a drop-in replacement for the sampler in an LTX IC-LoRA graph - put it exactly where SamplerCustomAdvanced would go:

LoadVideo → GetVideoComponents ──┐
EmptyLTXVLatentVideo ────────────┤
CLIPTextEncode (pos/neg) ────────┤→ LTXAddVideoICLoRAGuide ──► GearTiledFusionSampler ─► VAEDecodeTiled
                                 └─ (positive, negative, latent)                        ▲
                                                    GearSigmas ──────────────────────────┘

One rule the tooltip states bluntly: the latent input must be the latent returned by the IC-LoRA guide node. That latent carries the appended guide frames and the noise mask the sampler needs. Feed it a plain latent and the node raises - the guide's token grid won't match the canvas.

The inputs that matter

• tile_width / tile_height - the pixel window the IC-LoRA was trained at, in multiples of 32. 1024x576 is the default and a genuinely good start; match your IC-LoRA's training resolution if you know it.
• overlap_frac - how much each tile overlaps its neighbor. The author's own tooltip: "Never go below 0.5: a periodic grid appears." That's measured, not guessed - respect it. Lower overlap is faster, but the grid artifact is ugly.
• grid_cycle - shifts the tile grid every N steps to decorrelate the boundaries. 1 (fixed grid) is the cheapest and fine when VRAM-bound; 4 is the thorough setting and cures banding on motion.
• canvas_device - auto keeps the canvas on GPU while it fits, then falls back to CPU; force cpu for extreme canvases. VRAM scales with tile size, not canvas, so bigger/longer is mostly a time cost - the author has run 3840x2176x25 on a 96GB card.
• seed, cfg, sigmas - standard sampler fare. CFG defaults to 1, which is right for LTX distilled; feed sigmas from GearSigmas if you're using the pack's validated schedule.

The single output is latent - wire it into VAEDecodeTiled and you're done. For HDR work, feed the decoded frames into Gear · ACEScct Decode + Save EXR on the way out.

Installing

• ComfyUI Manager: search for ComfyUI_Gear, install, restart.
• Manual:

  cd ComfyUI/custom_nodes
  git clone https://github.com/oumad/ComfyUI_Gear
  cd ComfyUI_Gear
  pip install -r requirements.txt

  You'll also need the ComfyUI-LTXVideo pack for the IC-LoRA guide nodes this sampler depends on. Dependencies are light (numpy, Pillow, OpenEXR, opencv); nothing model-sized is downloaded by the pack.

Where people get burned

• latent_downscale_factor above 1 on the guide. The node raises on purpose - a downscaled guide's token grid won't line up with the tile crops. Re-attach the guide at factor 1.
• Batch size and dimensionality. It wants a single 5D video latent. Batch size 1 only - the code checks and throws otherwise.
• Overlap under 0.5. That periodic grid is real, and it's the #1 complaint-shaped footgun here. Keep 0.5+ and save your speed elsewhere.
• Massive canvas, small card. Leave canvas_device on auto and let it spill to CPU rather than forcing gpu and OOMing. It's slower, not broken.

This is a niche node for a specific workflow - but if your workflow is LTX IC-LoRA V2V above native resolution, it's the difference between "can't" and "runs overnight, seams-free."