Y7 Pad Image for Outpainting

If you've ever outpainted in ComfyUI and hit a "multiple of 8" error - or worse, got a subtly smeared canvas that you couldn't explain - this node is the fix. It's a clone of the built-in Pad Image for Outpainting, with exactly one thing added: a step input that snaps your padding values to the nearest multiple of 8 (or 16, or 64) before the pad happens. That's the whole trick, and it's a good one.

Why the snapping matters

Your image goes through a VAE, and VAEs downscale by 8x per side on the way into latent space (16 per side on newer models like Flux, which wants dimensions divisible by 64). If your padded canvas isn't a clean multiple, the latent gets rounded up and your "pad to 512" becomes "pad to 518" - and the outpaint result comes back offset, or with artifacts along the edge you were trying to extend. Manually doing (width + left + right) % 8 in your head before every outpainting run is exactly the kind of busywork nobody should do.

The built-in node assumes you'll type clean numbers. This one forgives you.

How it works

The source is short and honest: it fills a new, larger canvas with 0.5 grey (torch.ones * 0.5), drops your original image into the middle, and returns that plus a mask that's white in the padded region and feathered near the border. Each of left, top, right, and bottom (all INT, default 0) is snapped to the nearest multiple of step - ties round up - independently, so you can pad 64px top but 200px bottom and both land on the grid. feathering (default 40) is the width of the soft gradient where old and new meet; more feathering means the model blends the extension instead of seeing a hard seam.

Outputs are IMAGE and MASK. Wire the image and mask into a VAEEncodeForInpaint (or a fill model), set denoising around 0.8–1.0 for the extension area, and prompt for what should live in the new space. The mask is what protects your original pixels - outpainting is the one case where the mask really earns its keep when the original has to survive.

The gotchas

• Your typed value is not what gets applied. Snap the pad values and the node quietly changes them. Need a literal 256px? Set step to 1. That's the escape hatch, and it's the first thing to check if your output dimensions look slightly off.
• Feathering silently turns off if feathering * 2 isn't smaller than the image's own width and height - the math skips it rather than failing, so a tiny image can end up with a hard mask edge and a visible pasted-on look. Drop the feathering value, or pad more.
• This node only snaps; it won't make your composition obey the grid. Plan the pad, snap, then look at the actual output size before you sample.

Install

ComfyUI Manager → Custom Nodes Manager → search "Y7Nodes", install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/yushan777/ComfyUI-Y7Nodes
pip install -r requirements.txt

Note the pack's requirements.txt pulls in heavy stuff (transformers, sentencepiece, lmstudio) for its other nodes - this one only needs torch, which you already have. The author describes the whole pack as "probably only useful to me," but this is the rare quality-of-life clone that quietly saves every outpainting run a moment of arithmetic.