ApplyRAUNetSimple

ApplyRAUNet gets all the attention, but if you've never set up a U-Net block pairing in your life, this is the node you actually want. ApplyRAUNetSimple is the same HiDiffusion RAU-Net effect - downscale the latent early in generation so an SD1.5 model can plan composition at a resolution it understands, then restore full res for detail, so you can generate at 1536–2048px without doubled limbs and tiled wallpaper. The difference is it skips the block math entirely: you pick a resolution range, it picks the blocks and timings.

The author is characteristically blunt about this node's place in the world: it only helps when you're generating significantly above what the model supports, and it can actively hurt quality at low resolutions. It's a "use it because you have to" node, not a "free quality" button. If you're at native res, leave it out of the graph.

The three inputs

• model_type - auto, SD15, or SDXL. auto inspects the model and guesses; the tooltip's guidance is that SD15 also covers SD 1.4 and 2.x.
• res_mode - high (1536-2048), low (1024 or lower), or ultra (over 2048). This is a preset hint, not a resolution check - the node even says it doesn't need to match your actual canvas exactly. The fun note: choosing low with SDXL just disables RAUNet entirely, because SDXL already generates natively at 1024.
• upscale_mode / ca_upscale_mode - how the latent is restored in the upscale and cross-attention blocks. default maps to bicubic, which is the author's recommendation anyway. You can override to bislerp or whatever floats your boat.

Model in, patched MODEL out. That's the whole node. It logs exactly which preset it applied - block numbers, start/end percentages, upscale modes - so if you want reproducible generations you can copy those values into the advanced ApplyRAUNet and iterate from there.

For reference, the SD1.5 high preset is the classic setup: input 3 / output 8, active from 0.0 to 0.5, with the CA layer (input 1 / output 11) running 0.0 to 0.35. ultra stretches the effect longer to cope with 2048+. Those are reasonable defaults the author has tuned by trial and error - and he's honest that the presets may still shift as he finds better ones.

Install and caveats

ComfyUI Manager → search comfyui_jankhidiffusion → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/blepping/comfyui_jankhidiffusion

No model files, no extra dependencies - this pack ships as pure Python against the ComfyUI API.

All the advanced node's caveats apply here too: it's for SD1.x/2.x/SDXL only, SDXL is sensitive (prefer ancestral/SDE samplers with it), and it doesn't play nicely with ControlNet while the scaling effect is active - the node attempts a conditioning-rescale workaround, but shortening the effect window and using an ancestral sampler is the reliable path. Don't stack it with Deep Shrink, since they target the same blocks. If the block numbers ever do start bugging you, the advanced node is where that curiosity gets answered.