Perturbed-Attention Guidance(PAG) 笔记

Self-Rectifying Diffusion Sampling with Perturbed-Attention Guidance
Github

摘要

近期研究表明，扩散模型能够生成高质量样本，但其质量在很大程度上依赖于采样引导技术，如分类器引导（CG）和无分类器引导（CFG）。这些技术在无条件生成或诸如图像恢复等各种下游任务中往往并不适用。在本文中，我们提出了一种新颖的采样引导方法，称为Perturbed-Attention Guidance（PAG），它能在无条件和条件设置下提高扩散样本的质量，并且无需额外的训练或集成外部模块。PAG 旨在通过去噪过程逐步增强样本的结构。它通过用单位矩阵替换 UNet 中的self-attention map来生成结构退化的中间样本，这是考虑到自注意力机制捕捉结构信息的能力，并引导去噪过程远离这些退化样本。在 ADM 和 Stable Diffusion 中，PAG 在条件甚至无条件场景下都显著提高了样本质量。此外，在诸如空提示的 ControlNet 以及图像修复（如修补和去模糊）等现有引导（如 CG 或 CFG）无法充分利用的各种下游任务中，PAG 也显著提高了基线性能。
在这里插入图片描述
研究表明，在diffusion U-Net的self-attention 模块中，query-key 主要影响structure ，values主要影响appearance。

如果直接扰动Vt 的话，会导致 out-of-distribution (OOD)，因此选择使用单位矩阵替换query-key 部分。

那么具体扰动Unet的哪一部分呢？作者使用了5k个样本，在PAG guidance scale s = 2.5 and DDIM 25 step的条件下，表现最好的是mid-block “m0”
在这里插入图片描述

代码

Diffusers 已经支持PAG用在多种任务中，并且可以和ControlNet、 IP-Adapter 一起使用。

from diffusers import AutoPipelineForText2Image
from diffusers.utils import load_image
import torchpipeline = AutoPipelineForText2Image.from_pretrained("~/.cache/modelscope/hub/AI-ModelScope/stable-diffusion-xl-base-1___0",enable_pag=True,  ##addpag_applied_layers=["mid"], ##addtorch_dtype=torch.float16
)
pipeline.enable_model_cpu_offload()prompt = "an insect robot preparing a delicious meal, anime style"
generator = torch.Generator(device="cpu").manual_seed(0)
images = pipeline(prompt=prompt,num_inference_steps=25,guidance_scale=7.0,generator=generator,pag_scale=2.5,
).imagesimages[0].save("pag.jpg")

PAG代码细节

如果同时使用PAG和CFG，那么输入到Unet中prompt_embeds定义如下，也就是[uncond,cond,cond]

    def _prepare_perturbed_attention_guidance(self, cond, uncond, do_classifier_free_guidance):cond = torch.cat([cond] * 2, dim=0)if do_classifier_free_guidance:cond = torch.cat([uncond, cond], dim=0)return cond

PAGCFGIdentitySelfAttnProcessor2_0计算，其中[uncond,cond]正常计算SA，第二个cond则计算PSA。

class PAGCFGIdentitySelfAttnProcessor2_0:r"""Processor for implementing PAG using scaled dot-product attention (enabled by default if you're using PyTorch 2.0).PAG reference: https://arxiv.org/abs/2403.17377"""def __init__(self):if not hasattr(F, "scaled_dot_product_attention"):raise ImportError("PAGCFGIdentitySelfAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")def __call__(self,attn: Attention,hidden_states: torch.FloatTensor,encoder_hidden_states: Optional[torch.FloatTensor] = None,attention_mask: Optional[torch.FloatTensor] = None,temb: Optional[torch.FloatTensor] = None,) -> torch.Tensor:residual = hidden_statesif attn.spatial_norm is not None:hidden_states = attn.spatial_norm(hidden_states, temb)input_ndim = hidden_states.ndimif input_ndim == 4:batch_size, channel, height, width = hidden_states.shapehidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)# chunkhidden_states_uncond, hidden_states_org, hidden_states_ptb = hidden_states.chunk(3)hidden_states_org = torch.cat([hidden_states_uncond, hidden_states_org])# original pathbatch_size, sequence_length, _ = hidden_states_org.shapeif attention_mask is not None:attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)# scaled_dot_product_attention expects attention_mask shape to be# (batch, heads, source_length, target_length)attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])if attn.group_norm is not None:hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2)query = attn.to_q(hidden_states_org)key = attn.to_k(hidden_states_org)value = attn.to_v(hidden_states_org)inner_dim = key.shape[-1]head_dim = inner_dim // attn.headsquery = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)# the output of sdp = (batch, num_heads, seq_len, head_dim)# TODO: add support for attn.scale when we move to Torch 2.1hidden_states_org = F.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False)hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)hidden_states_org = hidden_states_org.to(query.dtype)# linear projhidden_states_org = attn.to_out[0](hidden_states_org)# dropouthidden_states_org = attn.to_out[1](hidden_states_org)if input_ndim == 4:hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width)# perturbed path (identity attention)batch_size, sequence_length, _ = hidden_states_ptb.shapeif attn.group_norm is not None:hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2)value = attn.to_v(hidden_states_ptb)hidden_states_ptb = valuehidden_states_ptb = hidden_states_ptb.to(query.dtype)# linear projhidden_states_ptb = attn.to_out[0](hidden_states_ptb)# dropouthidden_states_ptb = attn.to_out[1](hidden_states_ptb)if input_ndim == 4:hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width)# cathidden_states = torch.cat([hidden_states_org, hidden_states_ptb])if attn.residual_connection:hidden_states = hidden_states + residualhidden_states = hidden_states / attn.rescale_output_factorreturn hidden_states

经过Unet后，noise_pred的计算方法。

    def _apply_perturbed_attention_guidance(self, noise_pred, do_classifier_free_guidance, guidance_scale, t, return_pred_text=False):r"""Apply perturbed attention guidance to the noise prediction.Args:noise_pred (torch.Tensor): The noise prediction tensor.do_classifier_free_guidance (bool): Whether to apply classifier-free guidance.guidance_scale (float): The scale factor for the guidance term.t (int): The current time step.return_pred_text (bool): Whether to return the text noise prediction.Returns:Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]: The updated noise prediction tensor after applyingperturbed attention guidance and the text noise prediction."""pag_scale = self._get_pag_scale(t)if do_classifier_free_guidance:noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)noise_pred = (noise_pred_uncond+ guidance_scale * (noise_pred_text - noise_pred_uncond)+ pag_scale * (noise_pred_text - noise_pred_perturb))else:noise_pred_text, noise_pred_perturb = noise_pred.chunk(2)noise_pred = noise_pred_text + pag_scale * (noise_pred_text - noise_pred_perturb)if return_pred_text:return noise_pred, noise_pred_textreturn noise_pred