RAUNet

This one's a bit of an odd inclusion - it isn't about inpainting at all. RAUNet is this pack's implementation of HiDiffusion, a separate technique for getting a UNet-based diffusion model to generate cleanly at resolutions higher than it was trained for, without any retraining. The README frames it as an "internal rework to improve compatibility with other nodes" that landed in May 2024, and it's the third paper this pack builds on alongside BrushNet and PowerPaint. Nothing about it requires BrushNet or PowerPaint to be in your graph - you can drop it onto a plain model + KSampler workflow if higher-resolution generation is what you're after.

How it works

HiDiffusion's approach is to resize internal feature maps for part of the denoising process rather than running the whole UNet at your target resolution throughout - that's roughly how you get resolution gains without retraining the underlying model. This node exposes two separate step windows for two separate parts of that trick, straight from the README: du_start/du_end control when the "Downsample/Upsample resize" is active, and xa_start/xa_end control when the "CrossAttention resize" is active. Different parts of the network, different windows, both expressed as step ranges rather than a single on/off switch.

Inputs and outputs

• du_start (default 0) / du_end (default 4) - the step range where the downsample/upsample resize applies.
• xa_start (default 4) / xa_end (default 10) - the step range where the cross-attention resize applies.

All four are step indices (range 0–10000, same convention as BrushNet's start_at/end_at). Input is model; output is model - this is a pure model patch, wire it in place of your regular model before the sampler.

Installing it

Bundled with the whole pack, no separate install. ComfyUI Manager: search "BrushNet," install, restart. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/nullquant/ComfyUI-BrushNet && pip install -r requirements.txt. No model downloads needed for this node specifically - it patches whatever model you already have loaded.

Common issues

Don't combine it with other HiDiffusion implementations. The README explicitly lists blepping's jank HiDiffusion node as incompatible with this pack, for the same reason FreeU_Advanced is: both patch the same underlying UNet functions RAUNet needs, and stacking either one alongside it breaks things. If you're already running a HiDiffusion-style node from another pack, don't add this one on top.

Where it sits relative to BrushNet in a chain isn't spelled out. The README states elsewhere that model-modifying nodes should go before BrushNet, not after - a reasonable inference is that RAUNet follows the same rule if you're combining the two, but the README never actually demonstrates that combination, so treat it as untested territory rather than a documented workflow.

It's genuinely obscure. Zero recorded search traffic for this node on comfy.icu, and it doesn't come up in general community discussion the way BrushNet or PowerPaint do. That's not necessarily a quality signal - HiDiffusion-style resolution tricks are a niche technique to begin with - but don't expect much community troubleshooting to exist if you run into trouble. If results don't change or you hit an error, the likeliest culprit is a UNet-patching collision with something else in your graph, the same class of problem the pack's own incompatibility list warns about.