LTX Stage 1 Sampler MXD

LTX's distilled video models are fast, but the "fast" part only works if you use the exact sigma schedule the model was trained on - and that's where most people trip. This node is the Stage 1 sampler from Maxed Out's LTX workflow, and its headline feature is that the trained 8-step schedule is already baked in. You pick Distilled 8 Step and go. No copying sigma values out of a GitHub workflow, no "why does my LTX look washed out" posts.

The description on the node is worth quoting because it tells you the whole philosophy: "LTX-Video Stage 1 sampler for the distilled workflow. Use Distilled 8 Step for the trained schedule, or Custom Sigmas when intentionally testing a manual schedule." The "intentionally" is doing real work there - the default is right, and the manual mode is for people who know what they're testing.

How it works

In Distilled 8 Step mode it feeds a fixed, descending 9-sigma schedule (1.0 down to 0.0 - that's eight denoising steps) into a sampler with CFG 2.0 and euler_ancestral_cfg_pp by default, which is the pairing that matches how the distilled LTX model was trained. In Custom Sigmas mode it parses whatever comma/space/newline-separated values you paste into custom_sigmas - a descending schedule ending in 0.0 - and samples on that instead. Same sampler, your sigmas.

Inputs and outputs

Standard sampler inputs: model, positive, negative, latent_image, plus seed, cfg (default 2.0), and sampler_name (choose from euler_ancestral_cfg_pp, euler_cfg_pp, or euler). The two that define this node:

• mode - Distilled 8 Step or Custom Sigmas.
• custom_sigmas - only used in Custom Sigmas mode; multiline string of sigma values.
• ltx_preview - show live LTX video previews during sampling. Its tooltip has the gotcha: this downloads the taeltx VAE into your vae model folder the first time, if it's missing.

Output: latent, which feeds the Stage 2 refiner.

Installing it

Ships in the ComfyUI-MaxedOut pack. ComfyUI Manager search Maxed Out, or:

cd ComfyUI\custom_nodes
git clone https://github.com/Maxed-Out-99/ComfyUI-MaxedOut.git

Restart ComfyUI. The only thing it'll pull is that taeltx VAE for previews - a small file that lands in models/vae/ automatically.

Where people get burned

Don't set a custom schedule unless you're actually experimenting - the distilled model was trained on that fixed 8-step path, and deviating from it is how you get mushy results. Also keep CFG low (2.0 is the intended default); LTX distilled models aren't like SDXL, cranking CFG to 7 does real damage. And if previews are doing something weird, that's the taeltx VAE downloading or a preview toggle, not the sampling itself - the latent output is unaffected.