TCD Sampler 🍌

If you run Hyper-SD's "unified" LoRA - the ByteDance distillation that lets you dial step count from 1 to 8 at inference time instead of picking a fixed-step checkpoint - you've probably seen it paired with a sampler called TCD. This node is that sampler, on its own, as a SAMPLER output with exactly one control: gamma.

TCD stands for Trajectory Consistency Distillation, and it's the sampling side of Hyper-SD's story: Hyper-SD trains with Trajectory Segmented Consistency Distillation, and TCD sampling is what lets that same trained model be sampled at a flexible step count without a separate checkpoint per step target. The KB's own settings table for Hyper-SD lists the pairing directly: DDIM trailing for the fixed-step variant, or TCD for the unified one.

What gamma actually does

TCD's own paper introduces gamma as the sampler's stochasticity dial. At 0, sampling behaves close to a fully deterministic solver - cheap, but at very low step counts deterministic solvers can leave detail on the table. Turning gamma up trades some of that determinism for controlled randomness that can recover detail a pure ODE-style pass smooths away, at the cost of slightly more seed sensitivity. The default here, 0.3, sits in the range most Hyper-SD guides and the paper itself land on as the practical sweet spot - a reasonable starting point rather than something you need to tune from scratch.

That's the one honest gap here: the modidex knowledge base only mentions TCD as a table entry alongside Hyper-SD, without covering the paper or the gamma parameter directly - the explanation above comes from general knowledge of the technique rather than a KB citation, so treat it as a starting mental model, not gospel.

Using it

Wire the SAMPLER output into a KSampler (Advanced)'s sampler slot, or into SamplerCustomAdvanced if you're assembling the pipeline manually. This is a distilled-model sampler - it's built for Hyper-SD (or other TCD-trained) checkpoints and LoRAs running at low step counts and low CFG, not for driving a full 20–50 step pass on an ordinary checkpoint. Running normal-checkpoint settings through a distillation-specific sampler doesn't make things faster; it's just the wrong tool paired with the wrong model.

Installing it

Via ComfyUI Manager: search cgem156-ComfyUI, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/laksjdjf/cgem156-ComfyUI

Restart ComfyUI. No model downloads for the node itself - you still need the Hyper-SD checkpoint or LoRA separately, same as any distillation technique; TCD Sampler only supplies the sampling algorithm.

Common issues & troubleshooting

Grainy or inconsistent output across steps. Gamma is probably too high for your step count - pull it back toward 0 and see if it settles.

Flat, lifeless results. The opposite problem - nudge gamma up from 0.3 in small increments.

No speed benefit, or oversaturated/artifacted output. Check you're actually running a Hyper-SD (or other TCD-compatible) checkpoint/LoRA. Distilled students are trained against a specific noise schedule and step range; running a normal checkpoint's usual settings (high CFG, 20+ steps) through a distillation-tuned sampler is the classic way to get oversaturated, broken-looking results - the pairing of sampler and model matters as much as the sampler itself.

Comparing it against DDIM trailing and seeing different results. Expected - they're different sampling strategies for the same underlying distilled model, not two names for the same thing. If you're chasing a specific published Hyper-SD result, match the sampler the guide actually used before assuming something's broken.