HiDiffusion SD 2.1

Here's the honest framing up front: SD 2.1 is the model the community mostly walked away from. Stability's 2.0/2.1 line swapped in an OpenCLIP text encoder, which broke compatibility with the entire 1.5 LoRA and embedding library, and the training-data decisions didn't win friends either. The SD 1.5 crowd stayed put and never really came back. So this node - HiDiffusion applied to stabilityai/stable-diffusion-2-1-base - is a niche of a niche.

But if you're one of the people who does run 2.1, the niche has a point. 2.1 is a 768-native model (better than 1.5's 512, worse than SDXL's 1024), and its jewelry, photorealism, and depth-model strengths have their defenders. This node lets that model draw at 1024×1024 natively instead of needing a hires fix, which is the same trick the whole pack performs: Megvii's HiDiffusion patches the pipeline so it generates far past its training resolution without the duplicated-subjects meltdown.

How it works

First run downloads stabilityai/stable-diffusion-2-1-base from HuggingFace. Then it's the standard HiDiffusion recipe: RA-UNet swaps the down/upsamplers for resolution-aware versions (with a step-ratio schedule that kicks in harder the larger the canvas), and MSW-MSA replaces global attention with shifted-window attention to keep memory sane. The node hard-codes a DDIM scheduler at 50 steps - there's no steps input, same as the SD 1.5 node - so what you control is the prompt, guidance, eta, width, height, and seed.

The output is a single image_out (IMAGE) you can drop into SaveImage or pipe onward.

The inputs

• positive_prompt / negative_prompt - plain strings; the default negative is a reasonable baseline.
• guidance_scale - default 7.5, standard for DDIM-style generation.
• width / height - default 1024×1024, which is already a third bigger than 2.1's 768 native. Stick to multiples of 64.
• eta - a DDIM-only parameter; leave at 1.0.
• seed - optional.

Install

The usual. ComfyUI Manager → search "ComfyUI HiDiffusion", or:

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

Restart ComfyUI and let the ~5 GB model download complete on first run.

Gotchas

The pack-wide caveats apply here more than anywhere, because this is the least-tested node of the four: the README flatly warns the pack is under development with known crashes and suggests testing in a separate install. The requirements.txt was removed on purpose, so if the node errors, read the console and pip install the missing dependency yourself - diffusers, transformers, and xformers are all required, and xformers needs to line up with your CUDA/PyTorch.

Should you actually use it? If you already have a reason to be on 2.1 - you like the OpenCLIP conditioning, you need its specific look - sure, it's the only easy way to run HiDiffusion on it. If you're choosing between this and the SD 1.5 node in the same pack, pick 1.5. Same trick, bigger ecosystem behind it, and it's the node that actually gets maintained-adjacent attention.