LTX Stage 2 Refiner MXD

LTX is the speed tier of local video - where Wan takes minutes, LTX drafts in seconds - and the trade-off shows in the draft quality. The way the community gets both speed and quality is a two-stage flow: generate a fast draft at base resolution, then run a refinement pass at higher resolution. LTX Stage 2 Refiner MXD is the second half of that pipeline, and it's a rare example of a custom sampler that ships with the exact schedule the model's authors used.

What it is

It's a KSampler-style node built specifically for the stage-2 refine step of the LTX 2.3 two-stage distilled workflow. Feed it your model, conditioning, and the latent coming out of stage 1, and it runs a short, low-denoise pass that sharpens the draft into the final clip.

The default configuration is the headline feature. In Official Refine mode it uses the sigma schedule straight from Lightricks' own LTX-2.3_T2V_I2V_Two_Stage_Distilled.json example - [0.85, 0.725, 0.4219, 0.0] - with euler_cfg_pp and CFG 1.0. The 0.85 starting sigma is the denoise strength: it's an intentional partial refine, not a full re-generation. The comments in the source are explicit that only the starting sigma is meant to vary, and that's what Custom Sigmas mode exists for.

How the sigma schedule works

Sigma schedules are the distilled-model way of controlling how much noise gets added back in each step. For stage 2 you don't want to destroy the stage-1 draft - you want to clean it up at 2x resolution. Four steps from 0.85 down to 0 is a very gentle, very fast pass, which is exactly why this two-stage trick works.

In Custom Sigmas mode you type your own descending schedule (comma, space, or newline separated) into custom_sigmas. The node validates it strictly: at least two values, strictly descending, and must end in 0.0. Get any of that wrong and it tells you instead of silently producing garbage.

Inputs that matter

• mode - Official Refine (default) or Custom Sigmas.
• latent_image - the stage-1 output. This is a refiner; it consumes, not creates.
• cfg - default 1.0. LTX distilled runs at CFG 1; raising it is usually a mistake here.
• sampler_name - euler_cfg_pp (default), euler_ancestral_cfg_pp, or euler.
• seed - with control_after_generate wired in for reproducibility.
• ltx_preview - per the node's schema this toggles LTX previews and would download the taeltx VAE to your vae folder if missing. Note that recent versions of this pack moved live previews out to the standalone Live-Preview-MXD pack, so if the field doesn't appear, that's why - previews now attach automatically.

Install

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

Restart, or ComfyUI Manager → search "Maxed Out". No extra dependencies beyond ComfyUI core.

Troubleshooting

• "Custom Sigmas must be in descending order" / "must end with 0.0" - you're in Custom Sigmas mode and the schedule failed validation. Check your commas and your ordering.
• Output looks unchanged - a 0.85→0 refine is subtle by design. If you expected a big change, you're using stage 2 as a generator; it's not.
• Sampling a non-LTX model with this - the schedule is tuned for the LTX 2.3 distilled pair. It'll run on other models, badly. Keep it on the LTX two-stage workflow.

Paired with the pack's Stage 1 sampler, it turns the LTX speed→quality pipeline into a two-node drop-in instead of a sigma-engineering project.