LTX-2 One-Shot Low-RAM Looper ??

Most long-video strategies in ComfyUI are "generate a segment, stop, wire the next segment, run again." That's tedious, and it's where the node spaghetti creeps in. This node is the "just run it" option: one sampler node that takes your full latent, walks through it in temporal tiles, and returns the whole denoised result in a single execution - while aggressively managing VRAM and RAM so it works on hardware that can't hold a long LTX-2 sequence at once.

It's the heavyweight sibling of IAMCCS_LTX2_LoopingSampler. The LoopingSampler is a clean, self-contained tiler; this one is the kitchen sink - spatial tiling on top of temporal tiling, per-chunk cache cleanup, model unloading, conditioning image support, and source-lock controls. If your card chokes on even moderately long clips, this is the node the pack points you at.

How it works

It chunks the latent with temporal_tile_size (default 80) and temporal_overlap (default 24) in pixel frames, like the LoopingSampler. The differences:

• Spatial tiling too - horizontal_tiles / vertical_tiles (up to 4 each) split each frame spatially, with spatial_overlap for seam blending. This is how 12GB cards get through frames they couldn't touch otherwise.
• Aggressive memory management - unload_all_between_chunks (default on) and soft_empty_cache_between_chunks (default on) drop model state between tiles; cleanup_every_n_chunks (default 1) controls how often the full cleanup runs. This is the "low RAM" in the name, and it costs speed.
• Conditioning support - optional_cond_images with optional_cond_image_indices (which frames they attach to, default "0") at cond_image_strength, plus optional_guiding_latents for V2V-style guidance.
• Continuity knobs - guiding_strength, source_lock_strength (hold the source fixed, 0 = off), adain_factor (default 1.0, per-tile statistics alignment), and guiding_start_step / guiding_end_step to scope when guidance applies.

Important dependency: unlike the LoopingSampler, this node requires the official ComfyUI-LTXVideo pack - it resolves LTXVInContextSampler and LTXVExtendSampler from it at runtime and raises a clear error if they're missing.

Inputs and outputs

• Core: model, vae, noise, sampler, sigmas, guider, latents.
• Tiling: temporal_tile_size, temporal_overlap, horizontal_tiles, vertical_tiles, spatial_overlap.
• Guidance/memory: the full optional list above.

Outputs: denoised_output (LATENT) and a report STRING with the run's statistics.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

And make sure ComfyUI-LTXVideo is installed (Manager → search "ComfyUI-LTXVideo", or git clone https://github.com/Lightricks/ComfyUI-LTXVideo). Restart both. The README's baseline is ComfyUI ≥ 0.3.0, Python ≥ 3.12, Torch ≥ 2.8.

Gotchas

Memory toggles are speed killers, and the defaults are tuned for safety, not pace - if you have headroom, set unload_all_between_chunks: false and you'll feel the difference. The classic confusion is optional_cond_image_indices: it's a string like "0" or "0,120" listing which frames the cond images attach to, so a wrong value silently conditions the wrong part of the clip. And if it refuses to run with "Required LTX samplers were not found," that's not this pack being broken - ComfyUI-LTXVideo isn't loaded, full stop.