Get (y1, y2, x1, x2) bounding box from a binary mask. This mirrors the original InsertAnything implementation: when the mask is nearly empty,

Normalize bbox coordinates to [0,1] as in ObjectStitch.

Basic image -> tensor transform used by ObjectStitch (Stable Diffusion style). This is a minimal copy of the preprocessing from ObjectStitch's

Load a simple `x1 y1 x2 y2` bbox txt file.

Restore the parameters stored with the `store` method. Useful to validate the model with EMA parameters without affecting the

Cached in-memory ObjectStitch inference. This avoids re-loading weights for every call.

Save the current parameters for restoring later. Args: parameters: Iterable of `torch.nn.Parameter`; the parameters to be

Create sinusoidal timestep embeddings. :param timesteps: a 1-D Tensor of N indices, one per batch element. These may be

Create a 1D, 2D, or 3D average pooling module.

Convert a bbox to square, matching the original InsertAnything logic.

Paste the edited crop back into the original target image. This is a faithful port of the original InsertAnything geometry, including margin

Disable the OSInsert FluxFillPipeline patch and restore upstream behavior.

Enable the OSInsert FluxFillPipeline patch for the current Python process.

Expand image and mask by symmetric padding (original behavior).

Foreground (CLIP-style) image transform from ObjectStitch.

This matches the implementation in Denoising Diffusion Probabilistic Models: From Fairseq. Build sinusoidal embeddings. This matches

Load ObjectStitch model + sampler once. The returned model is moved to ``config.device`` and set to ``eval()``.