Feature Summary
supports lanpaint sampling in sd.cpp
Detailed Description
lanpaint uses Langevin dynamics approaches to do inpainting/outpainting with any model, without the need of a separately trained model. Instead of relying on ML approaches it models the conditioned stochastic processes of the (de)noising of inpaint image generation using the well-studied Langevin equations, thus effectively making normal i2i models consider mask contexts. IMO it is still the best AI-based image edit approach up to date that ensures pixel-to-pixel identity.
The core mechanism of this approach is within the sampler, we would need to port the sampler with the Langevin dynamics to sd.cpp to make this happen, I can start to work on a PR for this later this week.
Alternatives you considered
No response
Additional context
No response
Feature Summary
supports lanpaint sampling in sd.cpp
Detailed Description
lanpaint uses Langevin dynamics approaches to do inpainting/outpainting with any model, without the need of a separately trained model. Instead of relying on ML approaches it models the conditioned stochastic processes of the (de)noising of inpaint image generation using the well-studied Langevin equations, thus effectively making normal i2i models consider mask contexts. IMO it is still the best AI-based image edit approach up to date that ensures pixel-to-pixel identity.
The core mechanism of this approach is within the sampler, we would need to port the sampler with the Langevin dynamics to sd.cpp to make this happen, I can start to work on a PR for this later this week.
Alternatives you considered
No response
Additional context
No response