Make Inpaint Model

Dedicated inpainting checkpoints exist for a reason - they're trained with extra input channels so the model actually understands "this region is masked, fill it in coherently" instead of just doing a regular img2img pass and hoping the seams blend. The catch is you need a separate multi-gigabyte file for every base model you want that behavior on. This node skips that: point it at your custom SD1.5 checkpoint plus the two original Runway reference files, and it derives an inpainting version on the fly, in memory, without you having to hunt down or store a dedicated inpaint checkpoint for every model you use.

What it's actually doing

Back in October 2022, Runway shipped sd-v1-5-inpainting.ckpt alongside the base SD1.5 weights - a version of the UNet fine-tuned with five extra input channels (four for the masked-image latent, one for the mask) over 440k additional training steps. That's what gives a "real" inpainting checkpoint its edge: it was taught to treat the mask as structured input, not just noise to paint over.

This node takes the difference between that dedicated inpainting checkpoint and the plain SD1.5 base it was derived from, and applies that same delta onto whatever custom SD1.5 model you feed it. The result is a model that behaves like it was fine-tuned for inpainting, without you needing to actually store a separate multi-gigabyte inpainting variant of every anime, realism, or style checkpoint you own. It's a clever, cheap trick - and it only works because the two reference checkpoints it needs are both real, specific files, not arbitrary substitutes.

The inputs and output that matter

Three required inputs, and precision matters on two of them:

• model - the SD1.5 checkpoint you want to convert. This is the "give it inpainting powers" target.
• sd1_5_pruned - a dropdown of checkpoints in your models folder, specifically meant to be the original SD1.5 base weights: v1-5-pruned-emaonly.ckpt or v1-5-pruned-emaonly.safetensors.
• sd1_5_inpaint - same idea, but for sd-v1-5-inpainting.ckpt, the original Runway inpainting checkpoint.

Both of those need to be the actual original files, not "any SD1.5 checkpoint" - the node is computing a weight delta between them, so if either one isn't the genuine reference model, the math doesn't line up and you don't get a clean transplant.

Output is a single MODEL, ready to wire straight into a KSampler (or any detailer/inpainting node) exactly like a normal checkpoint's model output.

Where to actually get those two files

The README points at RunwayML's original HuggingFace repos - and here's the landmine: Runway deleted its entire HuggingFace org in August 2024, taking both of those files down with it. If you follow the README's links today, they're dead. The community re-hosted the weights within hours under the license's irrevocable copyright grant, and the canonical mirror now lives under the stable-diffusion-v1-5 org on HuggingFace rather than runwayml - search for stable-diffusion-v1-5/stable-diffusion-v1-5 and stable-diffusion-v1-5/stable-diffusion-inpainting. Grab both, drop them in your regular checkpoints folder, and they'll show up in the two dropdowns.

Installing it

Search ComfyUI-Hangover-Nodes in ComfyUI Manager, or clone it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Hangover3832/ComfyUI-Hangover-Nodes

Restart ComfyUI. No extra Python dependencies beyond what ComfyUI core already has - the node is doing checkpoint math, not running its own model.

Is this still worth using in 2026?

Honest take: SD1.5 dedicated inpainting has lost the quality argument. Community consensus by mid-2026 is that Qwen-Image-Edit or Flux 2 Klein, run through crop-and-stitch with a masked reference latent, simply produces better results on skin and hands than any SD1.5 inpainting checkpoint can. What SD1.5 hasn't lost is the speed argument - it's the reason people still reach for it on quick, VRAM-light inpainting passes, especially inside automated detailer workflows (ADetailer, Impact Pack) where you're running dozens of small masked fixes per image and don't want a 12B-parameter model in the loop for each one. If that's your use case - fast face/hand touch-ups on an SD1.5-based workflow - this node is still a genuinely useful shortcut. If you're chasing the best possible inpaint quality on a modern base model, this isn't that tool anymore; look at the newer edit-model-plus-mask approach instead.

Common issues

If both dropdowns show their placeholder text ("Please select the original SD 1.5 pruned model" / "...inpaint model") instead of a file list, it means ComfyUI hasn't found any checkpoints matching those slots - usually because the files aren't in your checkpoints folder yet, or you grabbed a re-uploaded/renamed variant rather than the exact original weights. Since this pack is marked unmaintained by its own author, don't expect fast fixes if a future ComfyUI core change breaks the checkpoint-diffing logic - but it's a small, self-contained node, so a stale install rarely drags anything else down with it.