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

This function returns the text prompts according to the requested `prompt_type` and `model`. Set `model` to "all" to return all the text prompts

in_channels: often obtained with `unet.config.in_channels`

This NMS processes boxes of all labels. It not only removes the box with the same label. Adapted from https://github.com/amusi/Non-Maxim

target_mask_shape: (h, w)

masks: numpy bool array

Computes the attend-and-excite loss using the maximum attention value for each token.

Computes the maximum attention value for each of the tokens we wish to alter.

fg_top_p, bg_top_p, fg_weight, and bg_weight are only used with max-based loss

This adds the ca loss with ref. Note that this should be used with ca loss without ref since it only enforces the mse of the normalized ca betwee

Each offset in `offset_list` is `(x_offset, y_offset)` (normalized).

in_channels: often obtained with `unet.config.in_channels`

This NMS processes boxes of each label individually.

v3 is equivalent to v2 but with new dictionary format for attention maps. The `saved_attn` is supposed to be passed to `save_attn_to_dict` in

Note: object_name needs to be a substring of each item in texts to make `count` and `get_box` in the predicate work

r""" Enable sliced attention computation. When this option is enabled, the attention module will split the input tensor in slices, to

r""" Parameters: `processor (`dict` of `AttentionProcessor` or `AttentionProcessor`): The instantiated processor c

`horizontal_shift_only`: only shift horizontally. If you use `offset` from `compose_latents_with_alignment` with `horizontal_shift_only=True`, th

attn_key: current key (a tuple of hierarchy index (up/mid/down, stage id, block id, sub-block id), sub block id should always be 0 in SD UNet