Inpainting VAE Encode

ComfyUI already ships a "VAE Encode (for Inpainting)" node, so the first question is why this one exists. The answer is one extra dial: mask_offset. This WAS version does the same job - encode an image plus a mask into a latent with a noise_mask attached - but lets you grow or shrink the mask before it ever touches the encoder, which is the difference between a clean inpaint and one with a visible seam.

How it works

The mechanic is standard masked-latent inpainting: the mask gets resized to match your image, applied to the pixels, and the whole thing gets VAE-encoded into a latent that carries a noise_mask alongside it. Whatever's under the white part of the mask is what your sampler is allowed to touch; the black part is locked. Where this node earns its keep is mask_offset - a positive value dilates (expands) the mask, a negative value erodes (shrinks) it, before encoding. Painting a mask that hugs the object too tightly tends to leave a visible ring where old and new content meet; nudging the offset a few pixels positive gives the model some buffer around the edge to blend into, which is exactly the boundary-artifact problem this node was built to reduce.

Worth knowing going in: encoding the masked region to latent noise like this generally wants a denoise of 1.0 downstream - you're regenerating the masked area from scratch, not lightly touching it up. If you want to inpaint at a lower denoise while keeping some of the original content as a hint, that's a job for InpaintModelConditioning instead, which core ComfyUI added specifically to handle the "keep original content, partial denoise" case that plain VAE-encode-for-inpaint can't do cleanly.

The inputs and outputs that matter

• pixels (IMAGE) - the image you're inpainting.
• vae (VAE) - the VAE to encode with; match it to whatever checkpoint you're sampling with.
• mask (MASK) - white = paint here, black = leave alone.
• mask_offset (default 6, range -128 to 128) - the one setting you'll actually tune. Positive dilates the mask (bigger buffer, softer blend, safer against leftover halos); negative erodes it (tighter to your painted region, less risk of eating into detail you wanted to keep). The default of 6 is a mild dilate, which is a reasonable starting point for most masks.

Single output: LATENT, ready to feed straight into a KSampler alongside your usual positive/negative conditioning.

How to install it

Via ComfyUI Manager: search "WAS_Extras", install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/WAS_Extras

Restart ComfyUI. No extra Python dependencies for this particular node - the pack's requirements.txt is only needed for the ksampler_sequence.py nodes (the CLIP Text Encode Sequence / KSampler Sequence family), not this one.

Common issues & troubleshooting

Results look muddy or washed out at the seam. That's the classic sign of too little mask buffer. Bump mask_offset up a few units before touching anything else - it's cheaper than fighting it with denoise or CFG.

The inpaint isn't taking at all, or looks like a no-op. Check your downstream KSampler's denoise. This node's noise_mask is built for a full 1.0 denoise on the masked region; a low denoise here will barely touch the mask and you'll wonder why nothing changed.

Fine detail you wanted to keep got erased. You probably over-dilated. Pull mask_offset toward zero, or negative, to tighten the mask back down to what you actually painted.

Need partial-denoise inpainting that respects the original pixels underneath. This node isn't the tool for that - reach for core ComfyUI's InpaintModelConditioning node instead, which was built precisely so you can inpaint at less than full denoise without the masked region collapsing to pure noise first.