Transforms standard latent layouts into packed multi-patch configurations.

Constructs and duplicates conditional multi-modal latents for execution batches.

Safely extracts tensors from VAE distributions.

Transforms structural noise output back to human-interpretable pixels.

Converts raw pixel space into initial VAE distributions.

Isolates valid computational vectors from padded hidden states.

Processes and extracts embeddings via the visual-language processor framework.

Calculates area-constrained ratios and standardizes spatial constraints.

Reverses multi-patch configurations back to classical latent layouts.

Performs pre-flight integrity checks on all pipeline parameters.

Calculates the sequence shift parameter based on sequence length ratios.

Create conversation template for image editing matching training format.

Resolves custom timesteps or sigmas from the given scheduler instance.

Generate or edit an image based on instruction. Args: instruction: Text instruction for generation/editing

Resizes a PIL image ensuring its dimensions are strict multiples of the given integer.

Prepare inputs for the model, handling understanding images.

Allocates and standardizes initial noise vectors and vision encodings for the step loop.

Process input images for understanding (editing).

Process input images for understanding (editing) using UND image VAE.

Set global random seed for reproducibility.

r""" Executes the multi-modal editing sequence based on the provided vision inputs and textual instructions. Args: images