LTX-2.5 KSampler (distilled) ⚡

Distilled models are picky, and LTX-2.5's are no exception: run them on the wrong step schedule and they don't look slightly worse, they look broken. This node is the antidote - a sampler with the two official LTX-2.5 schedules baked in as a two-option dropdown, wired through ComfyUI's real dual-CFG guider. There's no step-count widget to get wrong, no scheduler to misread, no chance to run an 8-step model at 30 steps and get artifacts instead of quality.

That "no wrong answers" property is the whole pitch. The stock schedulers (simple, karras, and friends) do not reproduce the trained distilled schedule, and a distilled model on the wrong schedule is the classic "is my install broken?" moment. This node passes the official ManualSigmas strings through verbatim - distilled (8 steps) for the stage-1 pass and refine (3 steps) for the final pass - using euler_ancestral as the sampler and a dual video/audio CFG guider where both streams run at CFG 1.0, exactly like the official workflow. CFG 1.0 is not a typo: the distilled bake is trained without classifier-free guidance, and cranking CFG on it is how you get oversaturated garbage. It's the same lesson as Lightning, Turbo, and every other distilled family - this node just makes it the default instead of a gotcha.

The inputs that matter

Required: model, positive, negative, latent_image, seed, schedule, sampler_name, video_cfg, audio_cfg. In practice:

• schedule - the one you'll flip. distilled (8 steps) for the first pass, then, after LTXV25LatentUpscale doubles the latent, run this same node again with refine (3 steps).
• video_cfg / audio_cfg - both default to 1.0 and both should stay there for the official recipe. They're separate because the guider is a true dual-CFG: video stream and audio stream each get their own value.
• seed - the usual. Different seeds are where LTX's character really lives; seed-hunting beats prompt-arguing on this model family.
• sampler_name - defaults to euler_ancestral and that's the one the schedule was tuned with. There are 44 samplers in the dropdown because it surfaces the full ComfyUI list, but you're not meant to shop here.

One LATENT out, which feeds Latent Upscale x2 (after distilled) or AV Decode (after refine).

Where it sits

The full chain, again, because every LTX-2.5 article ends up stating it: Models Loader → Img to Video → this node (distilled) → Latent Upscale x2 → this node (refine) → AV Decode. The v2v variant inserts Crop Video Guide between the distilled pass and the upscale.

Installing it

This is part of the ComfyUI-GGUF-Loader pack (ChrisColeTech's fork of city96's ComfyUI-GGUF). ComfyUI Manager: search "ComfyUI-GGUF-Loader". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
pip install --upgrade gguf

Restart.

Common issues

The most frequent mistake is not running the node twice - sampling once on distilled and calling it done. The refine pass is where the upscaled latent gets its final detail; skip it and you're leaving resolution on the table. Beyond that, the failure modes are mostly upstream: if the output looks like static, check the prep node's mode and whether images is connected, not the sampler settings. This node is the part of the chain that can't be the problem.