Img2ImgTurboEdgeLoader

This node has zero inputs and zero widgets, and it's still the part of the workflow people get stuck on. Img2ImgTurboEdgeLoader is the whole model-pipeline factory for the edge-to-image half of this pack: the first time you drag it in and run, it silently downloads a distilled, one-step model that turns Canny edge maps into finished images. The run node next to it is the dumb part. This is the heavy part.

What it actually loads

Under the hood it wraps CMU's img2img-turbo (Pix2Pix-Turbo) research code. The loader constructs a model from stabilityai/sd-turbo - the tokenizer, CLIP text encoder, VAE, and UNet all come from HuggingFace - then bolts on the edge_to_image LoRA-style adapter. That adapter ships as a single .pkl file fetched from CMU's servers:

https://www.cs.cmu.edu/~img2img-turbo/models/edge_to_image_loras.pkl

So the first run does a lot: a few GB of base model from HuggingFace, plus the edge LoRA. It looks hung, but it's just downloading. The .pkl lands in your ComfyUI/models/loras folder and the base model in the HuggingFace cache, so it's a one-time cost. After that the loader is instant.

The output is a single model output of type Img2ImgTurboEdgeModel. There's nothing to configure. It's a research wrapper, not a tunable loader.

Where it fits

Pair it with Img2ImgTurboEdgeRun. The workflow is: photo → CannyEdgePreprocessor → Img2ImgTurboEdgeRun, with this loader feeding the model input. One denoising step later you get the rendered image. It's the edge-to-image version of what a ControlNet canny setup does, except the speed is baked into the model itself rather than coming from a separate control network - no CFG, no sampler, no step count to get wrong.

How to install it

Through ComfyUI Manager, search ComfyUI-Img2Img-Turbo and install. Or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/chaojie/ComfyUI-Img2Img-Turbo
pip install -r ComfyUI-Img2Img-Turbo/requirements.txt

Then restart ComfyUI. The requirements are peft, diffusers==0.25.1, transformers==4.35.2, and xformers>=0.0.20.

Gotchas

• The diffusers pin is the pack's biggest failure point. The code relies on diffusers APIs that newer versions broke, and the README is upfront about it: if installing another node upgrades diffusers, put it back with pip install 'diffusers>=0.24.0,<=0.25.1'. That version pin is exactly the kind of cross-node dependency conflict the ComfyUI ecosystem is famous for.
• CUDA only. The code hard-calls .cuda(); there's no CPU or Apple Silicon fallback.
• If you delete the .pkl from your loras folder, the loader just re-downloads it. Harmless, just slow.

It's a thin node with a big first-run footprint. Budget a couple of minutes the first time, then it's forgettable.