Alright, let's get the unpopular opinion out of the way first: I'm starting to think the entire concept of SD inpainting as a "clean" tool is a bit of a collective hallucination. We've all seen the flawless YouTube demos where a stray object vanishes without a trace, but in the real, gritty trenches of my own workflows, it's a blurry, artifact-ridden mess nine times out of ten. The promised "seamless" blend is, more often than not, a sad, translucent smear that screams "I was edited."
My specific gripe—and the reason for this thread—is the persistent, halo-like blurry edge. You know the one. You mask out a lamppost, set the denoising to a hopeful 0.75, prompt for "clear sky," and what you get is a ghostly, out-of-focus aura where the lamppost used to be. It's not blended; it's just broken. The surrounding pixels seem to have a panic attack and refuse to commit to any coherent structure.
I've run the gauntlet of supposed fixes:
- The classic "just use the mask blur slider" advice, which only seems to soften the failure, not eliminate it.
- Toggling between "inpaint masked" and "inpaint not masked" like a frantic switchboard operator, each yielding its own unique flavor of visual dissonance.
- Cranked the padding pixel count up and down, which mostly just moves the blurry border inward or outward, like adjusting the width of a poorly poured concrete footer.
- Experimented with every combination of sampler and CFG under the sun, from the venerable Euler a to the supposedly magic DPM++ 2M Karras. The blur remains, a constant companion.
This leads me to a heretical line of questioning for this community. Is the core issue that we're expecting a simple, one-step fix for what is fundamentally a multi-layered compositing problem? Are we misusing the tool? Everyone talks about model choice and prompts, but the elephant in the room seems to be the native inpainting model's... let's call it "interpretive" approach to edge reconciliation.
I'm deeply curious about your under-the-hood experiences, especially if you've moved beyond the Automatic1111/GUI comfort zone into ComfyUI or scripting. Have you found a node workflow or a specific pre-process/post-process stack that actually banishes the blur? Or is the real pro move to just accept that SD inpainting is only good for generating *new* content in the hole, and any serious "clean" removal requires a round-trip to a proper image editor like Photoshop or GIMP before the image even touches the diffusion process?
Let's be brutally honest about the vendor value here. The feature is sold as integrated and intelligent, but the output often feels like a beta plugin. Prove me wrong, or perhaps, join me in a chorus of sardonic acceptance.
—Bella
Price ≠ value.