SamplerTCD EulerA

If you're using this pack at all, this is the node you're using. SamplerTCD Euler A is the finished sampler in dfl/comfyui-tcd-scheduler, the one the README points you to and the one wired into the pack's own example workflow. It's a modified Euler-ancestral step that implements the Trajectory Consistency Distillation update - the technique behind TCD speed LoRAs and Hyper-SD's unified LoRA. TCD was the 2024 answer to "I want LCM's speed but less mush," and a lot of people who tried both ended up preferring it in ComfyUI, once they found a scheduler that wasn't broken.

What it does and how

TCD's core idea is a timestep offset. Each step, the model is evaluated at the noise level you're actually at, but the update is aimed at an earlier point in the denoising trajectory - how much earlier is controlled by gamma. Concretely, the sampler converts the current sigma to a timestep t, targets down_t = (1 - gamma) * t, converts that back to a sigma (clamped so it never overshoots the next step), then runs a standard Euler-ancestral step toward that offset. The author's framing is a crossfade: at gamma 0 it's deterministic, at gamma 1 it's fully stochastic. Think of it as the distilled-model version of the Karras-versus-Euler tension.

The one input that matters

Just one: gamma (FLOAT, default 0.3, range 0–1, step 0.01). That's the whole dial, and it's the same parameter the LCMScheduler-style tools use to control per-step stochasticity.

• 0 - fully deterministic; same seed, same image, every time. Good for reproducibility, but you're giving up the sampler's creativity.
• 0.3 - the default, and a solid starting point.
• Higher - more stochasticity. The author recommends pushing gamma up as you raise the step count. The pack's shipped workflow runs 0.6 at 8 steps, which is a fine template.

The output is a SAMPLER, which plugs into the sampler input of a SamplerCustom node, next to the SIGMAS from the pack's TCDScheduler. That pairing is the whole setup - scheduler makes the noise schedule, this makes the step function.

Installing it

Same as the rest of the pack. ComfyUI Manager: search ComfyUI-TCD-scheduler, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/dfl/comfyui-tcd-scheduler

Restart ComfyUI and you're done. No pip dependencies, no model downloads - the pack is a thin wrapper over ComfyUI's own sampler machinery, so the only real "setup" is loading the TCD or Hyper-SD LoRA that gives the node a purpose.

Where people get burned

• Steps and CFG. This sampler expects the distilled-model playbook: ~4–10 steps matching your LoRA, CFG down at 1–2. Run it at CFG 7 with 25 steps and you'll swear the node is broken.
• Don't confuse it with SamplerTCD. The plain one is a WIP that behaves like DDIM and leaves a debug print in your console. If a workflow you downloaded uses SamplerTCD, swap it for this node - same gamma, same wiring.
• It's an iteration tool, not a final renderer. Distilled sampling trades detail for speed, so the classic pattern is: farm compositions fast with TCD, then re-render the winner with your full model. Treat this node as the fast lane and it's great; expect final-quality output and you'll be disappointed.