Replicate cjwbw/hyper-sdxl-1step-t2i

Hyper-SD is ByteDance's other distillation method (the same team that did SDXL-Lightning), and this is the extreme end of it: one-step text-to-image on SDXL. One denoising pass, a full image out. It's the "how fast can we possibly go" configuration, and this node runs it on Replicate's cloud through ComfyUI-Replicate.

Set expectations: one step is about speed and rapid iteration, not final-quality output. Any distillation loses some quality versus the full model, and one-step loses the most. Hyper-SD's signature trick is a unified LoRA that works across 1–8 steps, so where Lightning needs a specific checkpoint per step count, Hyper flexes - but at 1 step you're firmly in "seed farming and previews" territory, not "hero shot."

How it works

Where Lightning uses progressive distillation, Hyper-SD uses trajectory-segmented consistency distillation - it segments the denoising path and trains the model to be consistent within each segment, which is what lets one LoRA cover a range of step counts. As with every distilled model, the guidance is baked in, so it wants low CFG and its designed step count. This particular node is the 1-step text-to-image preset, so most of that is fixed for you; you mostly just prompt. The node calls Replicate and returns an IMAGE.

Inputs and outputs that matter

Output is an IMAGE. This node keeps the surface small:

• prompt / negative_prompt - your text.
• width / height (1024) - SDXL native.
• num_outputs (1) - batch up to 4.
• seed - fix it to reproduce a result; leave it random to farm compositions.
• output_format (webp default) / output_quality (80) - file output settings.

Notice there's no steps or CFG knob exposed here - that's deliberate, because this is the fixed 1-step preset. force_rerun forces a fresh generation on identical inputs.

How to install it

ComfyUI Manager: search ComfyUI-Replicate, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/replicate/comfyui-replicate
cd comfyui-replicate
pip install -r requirements.txt

Restart, and set your token first:

export REPLICATE_API_TOKEN="r8_************"; python main.py

Token: replicate.com/account/api-tokens.

Common issues

Clean image output - the well-supported case, no pack caveats to worry about.

The main "issue" isn't a bug, it's a mismatch of expectations: a 1-step image looks like a 1-step image. Soft details, less coherence on complex scenes, the occasional smeared area. That's the tradeoff you signed up for. If the quality isn't enough, the move the community actually uses is to farm with the fast model and then re-render the winners with a fuller model - generate a bunch here to find a good composition and seed, then run that seed through a higher-step model (even the 4-step Lightning node is a step up) for the final.

And the standing reminder for this whole pack: it's a cloud call, so a missing REPLICATE_API_TOKEN errors on the first run, and each generation bills on Replicate - a slightly odd trade for a model whose entire selling point is that it's cheap and fast to run locally.