HiDiffusion SDXL Turbo

SDXL Turbo is a weird model. Stability AI distilled it to 1–4 steps so it renders almost instantly, but the adversarial distillation hammered it down to 512×512 native output - half of what regular SDXL does. That's the entire trade: speed for resolution. This node says no to the trade and hands Turbo the same HiDiffusion patch the rest of this pack uses, so a 4-step generation can push to 1024 and beyond without falling apart.

It's the "turn your brain off" node of the pack. There's no model picker, no sampler menu, no negative prompt box - just a prompt, a resolution, and a couple of numbers. Everything else is decided for you.

How it works

The node pulls stabilityai/sdxl-turbo straight from HuggingFace (fp16 variant), applies RA-UNet + shifted-window attention with a Turbo-specific switching schedule baked in, and runs a single 4-step pass. Because Turbo is a distilled model, the usual rules flip: guidance_scale defaults to 0 - that's not a bug, distilled models want CFG disabled, and bumping it is how you wreck the output. And there's no negative prompt at all, which is normal for this family (Lightning behaves the same way). If you're coming from regular SDXL workflows, the settings here will look wrong on purpose.

Resolution defaults to 1024×1024, double what Turbo was really trained to do. That's the whole point of the node - it's not trying to be the fastest SDXL in town, it's trying to be the fast SDXL that doesn't tile and duplicate at 1024+.

The inputs

Sparse, which is the charm:

• positive_prompt - your only real creative control.
• inference_steps - default 4, range 1–99. Stay at 1–4; that's what Turbo knows.
• guidance_scale - leave at 0. Seriously.
• width / height - default 1024×1024, divisible-by-64 sizes are safest.
• seed - optional, for reproducible runs.

Output is a single image_out (IMAGE) to wire into SaveImage or any upscaler.

Install

Same as the rest of the pack. ComfyUI Manager, search "ComfyUI HiDiffusion", or:

cd ComfyUI/custom_nodes
git clone https://github.com/florestefano1975/ComfyUI-HiDiffusion

Restart ComfyUI. First run downloads the Turbo model from HuggingFace - a few gigabytes, so don't panic at the progress bar.

Gotchas worth knowing

The pack's README warns it's unfinished and that users have hit crashes; the requirements.txt was stripped out to dodge dependency fights. If the node errors on load, read the console and install the missing bits manually - diffusers, transformers, and xformers (all three are called unconditionally in the code). xformers has to match your CUDA build, which is the classic stumbling block.

Honest take: this is a curiosity. Turbo at 4 steps is fun and this makes it bigger than it was ever meant to be, but the quality ceiling is lower than the plain SDXL node in the same pack - you're trading the distillation's speed for its weakness. If you want big AND good, use HiDiffusionSDXL and wait the extra seconds. If you want big AND instant-ish, this is your toy.