r""" Returns: `dict` of attention processors: A dictionary containing all attention processors used in the model with

This computes the mass center of the mask. normalize: the coords range from 0 to 1 Reference: https://stackoverflow.com/a/66184125

The `saved_attn` is supposed to be passed to `save_attn_to_dict` in `cross_attention_kwargs` prior to computing ths loss. `AttnProcessor` wil

save_ind_in_filename: This adds a global index to the filename so that two calls to this function will not save to the same file and overwrite th

image should be a PIL object or numpy array with range 0 to 255

Expand overall bboxes from a 3d list to 2d list: Input: [[box 1 for phrase 1, box 2 for phrase 1], ...] Output: [box 1, box 2, ...]

Apply gaussian filter to input. Arguments: input (torch.Tensor): Input to apply gaussian filter on. Returns:

Args: hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`. When continuo

r""" Args: sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor timestep (`torch.FloatTen

The `bboxes` should be a list, rather than a list of lists (one box per phrase, we can have multiple duplicated phrases). batched: E

object_positions: object indices in text tokens return_cross_attn: should be deprecated. Use `return_saved_cross_attn` and the new format.

Note: the returned input latents are scaled by `scheduler.init_noise_sigma`

saved_attns: a list of saved_attn (list is across timesteps) moves to cpu by default

latents: encoded from the image, should not have noise (t = 0) returns inverted_latents for all time steps

amp_loss_scale: this scales the loss but will de-scale before applying for latents. This is to prevent overflow/underflow with amp, not to adjust

Keys: key = "CompVis/stable-diffusion-v1-4" key = "runwayml/stable-diffusion-v1-5" key = "stabilityai/stable-diffusion-2-1-base"

so_center_box: using centered box in single object generation so_horizontal_center_only: move to the center horizontally only align_with

spec: the spec for generation (see generate.py for how to construct a spec) bg_seed: background seed overall_prompt_override: use custom

so_center_box: using centered box in single object generation so_horizontal_center_only: move to the center horizontally only align_with

so_center_box: using centered box in single object generation so_horizontal_center_only: move to the center horizontally only align_with

Disables custom attention processors and sets the default attention implementation.

base_w and base_h: make sure the shift is aligned in the latent and multiple levels of cross attention

Visualize cross attention for one token, across timesteps: `stage_id`th downsampling block, mean over all timesteps starting from step start, `bl

Visualize cross attention: `stage_id`th downsampling block, mean over all timesteps starting from step start, `block_id`th Transformer block, sec