LTX-2 Init Latent Sampler ??

Sometimes the loopers are overkill. If your long-video pipeline already handles the segment splitting elsewhere - say, with the queue/segment machinery in this same pack - then the sampler's job is just to denoise one segment's latent once, cleanly, without re-tiling it into pieces. That's this node: a deliberately simple single-pass sampler that takes the guider/sampler/sigmas your graph already built and runs them against the provided initialization latents.

The name is the intent. It samples the initialization latents you give it - the segment's starting state, which in the IAMCCS pipeline usually carries latent-bridge metadata from IAMCCS_LTX2_LoadLatentBridge - and hands back the denoised result. It exists to replace external looping samplers in graphs where the queue segmentation is already handling the long-video split, so you don't end up double-tiling.

How it works

Two paths, chosen automatically:

• Continuation path - if the latent carries a leading overlap (via iamccs_prev_latents or an inferred overlap) and you've passed optional_guiding_latents (the source-video latents in V2V work), it routes through LTX's LTXVExtendSampler to continue the previous state, carrying the seed offset and reference latent along. This is the path that keeps faces and structure from drifting in detailer/upscale passes.
• Simple path - otherwise it just denoises the latents once with the upstream guider/sampler/sigmas, no tiling, no extension.

The optional inputs (optional_cond_images, optional_positive_conditionings, optional_negative_index_latents, optional_normalizing_latents) pass through to the chosen sampler when present.

Dependency note: the continuation path needs ComfyUI-LTXVideo loaded (it resolves LTXVExtendSampler / LTXVInContextSampler from it). Without it, you get the simple path - and for detailer work that's exactly the case the source comments warn about: ignoring the guiding latents "turns the pass into a fresh generation," which is how faces and body structure drift even though resolution goes up.

Inputs and outputs

• Core: model, vae, noise, sampler, sigmas, guider, latents.
• Optional: optional_cond_images, optional_guiding_latents, optional_positive_conditionings, optional_negative_index_latents, optional_normalizing_latents.

Output: denoised_output (LATENT).

Install

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

With ComfyUI-LTXVideo installed alongside if you want the continuation path (it's what makes this node useful for segment detailers). Restart both.

Gotchas

The trap is using it for segment detailer/upscale passes without feeding optional_guiding_latents. The node will happily run the simple path and produce a sharper-looking result that no longer looks like the subject - the source code says this explicitly. If you're here for detailer duty, wire the source-video latents in. And if you get "SamplerCustomAdvanced not available," update ComfyUI; that's the one hard requirement for even the simple path.