[Feature Request] a Good Inpainting Model Just Dropped for adding
Hey, just wanted to share something really cool I came across — it's a new model called PixelHacker for image inpainting. But it’s not your typical filler—it actually understands structure and semantics when reconstructing missing parts of an image.
What makes it stand out is this clever idea: before the diffusion step, it splits the image into foreground and background and encodes them separately. Then it uses a latent-category guidance system to inject that info during denoising. The results are seriously clean and coherent—way more believable than most models I’ve seen.
There’s no web demo yet, but they’ve got tons of examples on their site, and they just open-sourced the whole thing—weights, code, everything. Super easy to plug into a project. If you’re doing anything that needs high-quality image recovery, smart edits, or generative visual content, I think this is absolutely worth integrating.
Source: Sanster/IOPaint