_summary_ Args: sample (torch.FloatTensor): _description_ timestep (Union[torch.Tensor, float, int]): _description_

fuse referencenet emb b c t2 h2 w2 into unet latents b c t1 h1 w1 with attn refer to musev/models/attention_processor.py::NonParamT2ISelfRefe

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

The [`Transformer2DModel`] forward method. Args: hidden_states (`torch.LongTensor` of shape `(batch size, num latent pix

r""" The [`UNet2DConditionModel`] forward method. Args: sample (`torch.FloatTensor`): The noisy input ten

_summary_ Args: unet_params (Dict): _description_ encoder_hidden_states (torch.Tensor): b t n d reference

generate b (n expand) (n expand) mask， where value of diag (0<=<=n_near) and first column of shape mat (n n) is set as big_value, others as s

covert video directory into tasks Args: path (str): _description_ output_dir (str): _description_ exts (Tuple[str]): _des

limite generation video frames numer to avoid gpu memory overflow

limite generation video frames numer to avoid gpu memory overflow

limite generation video frames numer 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 不通过models.unet_3d_condition.py:UNet3DConditionModel 定义、训练的模型 model is not defi

Decorator to register a function or class.

Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and Sample Steps are Flawed](http

sample sub_tensor Args: tensor (Union[torch.Tensor, np.ndarray]): _description_ n_sample (int): _description_ sample_rate

Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. Args: sample (`torch.FloatTens

Disables custom attention processors and sets the default attention implementation.

Sets the timesteps used for the diffusion chain. Supporting function to be run before inference. Args: num_inference_ste

Returns the image processor and clip image encoder for the img2img unclip pipeline. We currently know of two types of stable unclip models w

Returns the noising components for the img2img and txt2img unclip pipelines. Converts the stability noise augmentor into 1. a `StableUnC

Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion process from the learned m

Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion process from the learned m

Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion process from the learned m

Step function propagating the sample with the multistep DPM-Solver. Args: model_output (`torch.FloatTensor`): direct out

Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion process from the learned m

Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion process from the learned m

Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion process from the learned m

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

use independent ip_adapter attn 中的 to_k, to_v in unet ip_adapter： like ['1.to_k_ip.weight', '1.to_v_ip.weight', '3.to_k_ip.weight']的字典 Args: