TS Smart Inpaint

Mask-based inpainting has quietly become the reliable old tool of the editing stack - while the new instruction-edit models (Qwen-Edit, Kontext) happily redraw the whole frame and drift every pixel you didn't ask about, a masked inpaint keeps the untouched areas bit-exact. The catch has always been the plumbing: crop the masked region, upscale it so the model has resolution to spend, encode, sample, and feather the result back. TSSmartInpaint is that whole pipeline in a single node - you feed the full image, the mask, and a checkpoint, and it returns the finished edit.

How it works

It runs the classic crop-and-composite loop internally, which the README spells out: it crops the masked region plus a context band, upscales the crop toward a megapixel budget so a small selection gets full-res detail, VAE-encodes, samples, then feather-composites and latent-blends the result back so untouched pixels stay bit-exact. You never build the crop → sampler → stitch chain yourself.

The replace toggle picks which of two very different jobs it does:

• Replace (default) - full regeneration of the masked area, "Kontext-style": the crop becomes the reference_latents, and denoise is locked to 1.0. There's even an optional reference image input - wire one in and the hole is filled toward that picture's content ("fill with THIS").
• Refine - an ADetailer-style partial-denoise pass at the denoise value, no reference. This is the "polish the existing content" mode, like a face-repair pass.

The defaults are tuned for transformer-era models: steps 4, cfg 1.0, scheduler "simple" - which is a very Flux-flavoured starting point, so if you're on SDXL you'll likely want more steps and a real CFG. megapixels (1.5 default) is the processing budget: small crops upscale toward it, oversized crops downscale to it so a huge mask in an 8K frame can't hang the machine.

Two settings are proportional rather than fixed pixels, which is the right call: context_pct (8) is the surrounding context as a percent of the mask's own size, and feather_pct (3) is the edge blend, also relative - so a thin stroke doesn't get an over-soft ghost.

Wiring it

Inputs you'll definitely connect: model, vae, positive, negative, image, mask. The positive/negative conditioning describes what to generate inside the mask. The one output is the composite image.

Install

No extra dependencies - it's pure comfyui-timesaver.

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt

Restart, then search "TS Smart Inpaint".

Common issues

• "The whole image changed." You probably expected Replace to leave everything else alone. It does - but if your mask is large and context_pct is high, the context band gets regenerated too. Keep the mask tight.
• "Result looks mushy." A small crop got upscaled toward megapixels and the model had to invent detail. Raise megapixels for sharper detail, or lower it if it's hallucinating.
• "Denoise doesn't do anything." That's Replace mode - it's deliberately locked to 1.0. Switch to Refine for a partial pass.
• Visible seam. Raise feather_pct a little; 0 is a hard edge and that's almost never what you want on a real edit.

The honest trade: a one-node inpaint can't match a hand-built chain where you tune the sampler for the specific model - but it removes about nine nodes of scaffolding, and for 95% of "remove this, fix that" mask work it's the faster road. The KB's verdict on masks still applies: bit-identical pixels outside the mask are the one thing no instruction-editor gives you.