SamplerTCD

Let's get the honest answer out first: you don't want this node. SamplerTCD is the cautionary tale in the dfl/comfyui-tcd-scheduler pack, and the author says so in the README - it's a work-in-progress that currently just operates as DDIM with no gamma parameter. The finished sampler in the pack is SamplerTCD Euler A. Use that one.

This page exists because old workflows floating around the internet still reference SamplerTCD, and if one loaded into your graph, you deserve to know what you're looking at.

What it is

It's the sampler half of the pack's TCD implementation (Trajectory Consistency Distillation - the technique behind TCD and Hyper-SD speed LoRAs). It takes one input, gamma (FLOAT, default 0.3, range 0–1, step 0.01), and outputs a SAMPLER for ComfyUI's custom sampler nodes. In theory it's a modified DDPM step with TCD's timestep-offset trick. In practice, the offset logic is commented out in the source, there's a leftover print of sigma_from/sigma_to/sigma_to_s that dumps noise into your console, and gamma does effectively nothing. That's what "WIP" means in a repo that hasn't been touched since.

Why it's broken

TCD's core move is evaluating the model at the current timestep but aiming the update at an earlier one, with gamma controlling how far back. The working Euler A sibling does this correctly. SamplerTCD attempts the DDPM-flavored version and never got finished - the meaningful lines are # commented out. So you get a plain DDIM step wearing a TCD name, which means a TCD-trained LoRA gets fed updates it wasn't trained against. It'll produce images, sure. It just won't produce what the technique promises.

What to do if it's in your workflow

Swap it out:

SamplerTCD  →  SamplerTCD Euler A

Same gamma input, same SAMPLER output, same wiring into your SamplerCustom. You'll need the pack installed either way. And while you're at it, that console spam you may have blamed on ComfyUI - the sigma print - disappears with the swap too.

One sentence of due credit: this half-finished node is exactly how the pack's author credited @laksjdjf for the sigma-to-timestep conversion that made the Euler A version work. The WIP got superseded fast, which is the best outcome a WIP can hope for.