key

沿着 dim 纬度， 补齐 src 的长度到目标 target_length。 当 src 长度不如 target_length 时， 取其中 前 n_src_base_length 然后 repeat 到 target_length align length of src to

This does the final conversion step: take locally converted weights and apply a global renaming to them. It splits attention layers, and take

_summary_ Args: tensor (torch.Tensor): b c t h w src_index (torch.LongTensor): _description_ dst_index (torch.LongTensor)

"Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the prediction of x_0 at t

_summary_ Args: tensor (torch.Tensor): D1*D2*D3*D4... index (torch.LongTensor): D1*N or N, N<= tensor.shape[dim] dim (int

Updates paths inside resnets to the new naming scheme (local renaming)

Removes segments. Positive values shave the first segments, negative shave the last segments.

通过模型名字 初始化 referencenet，载入预训练参数， 如希望后续通过简单名字就可以使用预训练模型，需要在这里完成定义 init referencenet with model_name. if you want to use pretrai

通过模型名字 初始化Unet，载入预训练参数. init unet with model_name. 如希望后续通过简单名字就可以使用预训练模型，需要在这里完成定义 if you want to use pretrained model with simple nam

Updates paths inside resnets to the new naming scheme (local renaming)

存储tensor视频为gif、mp4等 Args: videos (Union[torch.Tensor, np.ndarray]): 五维视频tensor， 如 b c t h w，值范围[0-1] path (str): 视频存储路径，后缀会影响存储方式

fuse overstep tensor with weight of src into dst out = src_fused_part * weight + dst * (1-weight) for overlap Args: src (torch.Tensor

Updates paths inside attentions to the new naming scheme (local renaming)

Ensures interchangeability with schedulers that need to scale the denoising model input depending on the current timestep. A

Args: src (torch.Tensor): b c t h w dst (torch.Tensor): b c t h w

_summary_ Args: tensor (torch.Tensor): b*c*h dim (int): index (torch.LongTensor): b*d, source (torch.Tensor):

_summary_ Args: tensor (torch.Tensor): b*c*h dim (int): index (torch.LongTensor): b*d, value (torch.Tensor): b

_summary_ Args: data1 (torch.Tensor): b1*c1*h1*w1*... data1_index (torch.LongTensor): N, if dim=1, N=c1 data2 (torch.Tens

Takes a state dict and a config, and returns a converted checkpoint.

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

r""" Tokenize a list of prompts and return its tokens with weights of each token. No padding, starting or ending token is included.

When the length of tokens is a multiple of the capacity of the text encoder, it should be split into chunks and sent to the text encoder indi

limite generation video shape to avoid gpu memory overflow

limite generation video shape to avoid gpu memory overflow

limite generation video shape to avoid gpu memory overflow

通过模型名字 初始化Unet，载入预训练参数. init unet with model_name. 该部分都是通过 models.unet_3d_condition.py:UNet3DConditionModel 定义、训练的模型 model is defined and trai

r""" Pad the tokens (with starting and ending tokens) and weights (with 1.0) to max_length.

Updates paths inside attentions to the new naming scheme (local renaming)

_summary_ Args: emb (torch.Tensor): b ( h w) d height (int): _description_ width (int): _description_ scale (floa

generate long video with end2end mode 1. prepare vision condition image by assingning, redraw, or generation with text2image module w

类似controlnet text2img pipeline。 输入视频，用视频得到controlnet condition。 目前仅支持time_size == step，overlap=0 输出视频长度=输入视频长度 simil

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

使用model_path更新pipeline中的基础参数 Args: pipeline (DiffusionPipeline): _description_ model_path (str): _description_ text_sd_mo

使用 lora 更新pipeline中的unet相关参数 Args: pipeline (DiffusionPipeline): _description_ lora_dict (Dict[str, Dict]): _description_

use independent ip_adapter attn 中的 to_k, to_v in unet ip_adapter： dict whose keys are ['1.to_k_ip.weight', '1.to_v_ip.weight', '3.to_k_ip.weight'

更新T2V模型中的T2I参数. update t2i parameters in t2v model Args: unet (nn.Module): _description_ sd_model (Tuple[str, nn.Module]): _descr

Args: funcs (list): 继承`BaseAttribute2Text`并实现了`__call__`函数的类. Inherited `BaseAttribute2Text` and implemented the `__call__` funct

纯空模板，无论输入啥，都只返回空格字符串作为prompt。 Args: template (str): 符合只输出空格字符串的模板， name (str, optional): 该模板字符串名字，暂无实际用处. Defaults to

Constructs the noise schedule of Karras et al. (2022).

Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of (1-beta) over time from t = [

Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of (1-beta) over time from t = [

Convert the model output to the corresponding type that the algorithm (DPM-Solver / DPM-Solver++) needs. DPM-Solver is designed to d

Creates a config for the diffusers based on the config of the LDM model.

One step for the first-order DPM-Solver (equivalent to DDIM). See https://arxiv.org/abs/2206.00927 for the detailed derivation.