HiDiffusion SD 1.5

SD 1.5 was trained at 512×512, and the moment you push past that you meet its worst habit: duplicated subjects. Ask for one mountain at 768px and the model helpfully gives you two. The classic fix is a two-pass hires fix - generate small, upscale, denoise again - which is slow and fiddly. This node tries something bolder: it applies Megvii's HiDiffusion patch to the raw SD 1.5 pipeline so the model can generate at 1024×1024 (or bigger) in one pass, natively.

That's the pitch, and honestly it's a fun one. SD 1.5 in 2026 is a cheap idea generator more than a primary tool - the community's own words are "I use SD1.5 when I want prompt rebellion" - and this node hands that scrappy creativity a much bigger canvas than it was ever allowed.

How it works

The node downloads runwayml/stable-diffusion-v1-5 from HuggingFace on first run, then patches it two ways. RA-UNet replaces the UNet's down/upsamplers with resolution-aware versions that adapt to canvas size, which is what stops the doubled-subject meltdown; a switching schedule kicks that in partway through denoising (more aggressive the bigger you go). MSW-MSA swaps global attention for shifted-window attention to keep memory and time from exploding. The result is a real resolution bump instead of an upscale - it draws new detail, it doesn't invent it from a blur.

Note what you don't get: no steps control. The node is fixed to a DDIM scheduler at its pipeline default of 50 steps. You set the prompt, guidance, eta, width/height, and seed. That's the whole panel.

The inputs

• positive_prompt / negative_prompt - the defaults are decent starting points, but they're just strings.
• guidance_scale - default 7.5; the classic 1.5 sweet spot.
• width / height - default 1024×1024. For SD 1.5 that's already double native; the T1 ratio in the patch gets stiffer at 2048, so the model self-corrects harder the further you stretch it.
• eta - DDIM-only parameter, leave at 1.0 unless you know why you're touching it.
• seed - optional.

Output is one image_out (IMAGE), which you can feed into SaveImage or an upscaler.

Install

ComfyUI Manager, search "ComfyUI HiDiffusion", or the manual route:

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

Restart ComfyUI. The first run downloads the SD 1.5 weights (~4 GB) from HuggingFace, so give it a minute.

Where people get burned

Same story as the rest of this pack: the README flags it as under development with known crashes, and requirements.txt was deliberately removed to avoid dependency conflicts. If the node won't load, the console will tell you what's missing - the code unconditionally needs diffusers, transformers, and xformers, and xformers must match your CUDA/PyTorch combo.

Manage expectations beyond that. This is a 2024-packaged research toy, unmaintained since, and the whole pack's reddit buzz came and went in about two months. It works on a current ComfyUI, but if you actually need big SD 1.5 output for real work, the standard hires-fix workflow is still the dependable choice. Reach for this when you want to see the trick.