diffusers-Load pipelines,models,and schedulers

https://huggingface.co/docs/diffusers/using-diffusers/loadinghttps://huggingface.co/docs/diffusers/using-diffusers/loading

有一种简便的方法用于推理是至关重要的。扩散系统通常由多个组件组成，如parameterized model、tokenizers和schedulers，它们以复杂的方式进行交互。这就是为什么我们设计了DiffusionPipeline，将整个扩散系统的复杂性包装成易于使用的API，同时保持足够的灵活性，以适应其他用例，例如将每个组件单独加载作为构建块来组装自己的扩散系统。

1.Diffusion Pipeline

DiffusionPipeline是扩散模型最简单最通用的方法。

from diffusers import DiffusionPipelinerepo_id = "runwayml/stable-diffusion-v1-5"
pipe = DiffusionPipeline.from_pretrained(repo_id, use_safetensors=True)

也可以使用特定的pipeline

from diffusers import StableDiffusionPipelinerepo_id = "runwayml/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(repo_id, use_safetensors=True)

Community pipelines是原始实现不同于DiffusionPipeline，例如StableDiffusionControlNetPipeline.

1.1 local pipeline

from diffusers import DiffusionPipelinerepo_id = "./stable-diffusion-v1-5" # local path
stable_diffusion = DiffusionPipeline.from_pretrained(repo_id, use_safetensors=True)

from_pretrained()方法在检测到本地路径时不会下载。

1.2 swap components in a pipeline

可以使用另一个兼容的组件来自定义任何流程的默认组件。定制非常重要，因为：

更改调度器对于探索生成速度和质量之间的权衡是重要的。
模型的不同组件通常是独立训练的，您可以用性能更好的组件替换掉现有组件。
在微调过程中，通常只有一些组件（如UNet或文本编码器）进行训练。

from diffusers import DiffusionPipelinerepo_id = "runwayml/stable-diffusion-v1-5"
stable_diffusion = DiffusionPipeline.from_pretrained(repo_id, use_safetensors=True)
stable_diffusion.scheduler.compatibles

from diffusers import DiffusionPipeline, EulerDiscreteScheduler, DPMSolverMultistepSchedulerrepo_id = "runwayml/stable-diffusion-v1-5"
scheduler = EulerDiscreteScheduler.from_pretrained(repo_id, subfolder="scheduler")
stable_diffusion = DiffusionPipeline.from_pretrained(repo_id, scheduler=scheduler, use_safetensors=True)

可以将PNDMScheduler更换为EulerDiscreteScheduler，在回传到DiffusionPipeline中。

1.3 safety checker

safety checker可以根据已知的NSFW内容检查生成的输出，

from diffusers import DiffusionPipelinerepo_id = "runwayml/stable-diffusion-v1-5"
stable_diffusion = DiffusionPipeline.from_pretrained(repo_id, safety_checker=None, use_safetensors=True)

1.4 reuse components across pipelines

可以在多个pipeline中可以重复使用相同的组件，以避免将权重加载到RAM中2次

from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipelinemodel_id = "runwayml/stable-diffusion-v1-5"
stable_diffusion_txt2img = StableDiffusionPipeline.from_pretrained(model_id, use_safetensors=True)components = stable_diffusion_txt2img.components

可以将components传递到另一个pipeline中，无需将权重重新加载到RAM中：

stable_diffusion_img2img = StableDiffusionImg2ImgPipeline(**components)

下面的方式更加灵活：

from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipelinemodel_id = "runwayml/stable-diffusion-v1-5"
stable_diffusion_txt2img = StableDiffusionPipeline.from_pretrained(model_id, use_safetensors=True)
stable_diffusion_img2img = StableDiffusionImg2ImgPipeline(vae=stable_diffusion_txt2img.vae,text_encoder=stable_diffusion_txt2img.text_encoder,tokenizer=stable_diffusion_txt2img.tokenizer,unet=stable_diffusion_txt2img.unet,scheduler=stable_diffusion_txt2img.scheduler,safety_checker=None,feature_extractor=None,requires_safety_checker=False,
)

1.5 checkpoint variants

以torch.float16保存，节省一半的内存，但是无法训练，EMA不用于推理，用于微调模型。

2. models

from diffusers import UNet2DConditionModelrepo_id = "runwayml/stable-diffusion-v1-5"
model = UNet2DConditionModel.from_pretrained(repo_id, subfolder="unet", use_safetensors=True)

所有的权重都存储在一个safetensors中，可以用.from_single_file()来加载模型。safetensors安全且加载速度快。

2.1 load different stable diffusion formats

.ckpt也可以用from_single_file()，但最好转成hf格式，可以使用diffusers官方提供的服务转：https://huggingface.co/spaces/diffusers/sd-to-diffusers

也可以使用脚本转：https://github.com/huggingface/diffusers/blob/main/scripts/convert_original_stable_diffusion_to_diffusers.py

python ../diffusers/scripts/convert_original_stable_diffusion_to_diffusers.py --checkpoint_path temporalnetv3.ckpt --original_config_file cldm_v15.yaml --dump_path ./ --controlnet

A1111 Lora文件，diffusers可以使用load_lora_weights()加载lora模型：

from diffusers import DiffusionPipeline, UniPCMultistepScheduler
import torchpipeline = DiffusionPipeline.from_pretrained("andite/anything-v4.0", torch_dtype=torch.float16, safety_checker=None
).to("cuda")
pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config)# uncomment to download the safetensor weights
#!wget https://civitai.com/api/download/models/19998 -O howls_moving_castle.safetensorspipeline.load_lora_weights(".", weight_name="howls_moving_castle.safetensors")prompt = "masterpiece, illustration, ultra-detailed, cityscape, san francisco, golden gate bridge, california, bay area, in the snow, beautiful detailed starry sky"
negative_prompt = "lowres, cropped, worst quality, low quality, normal quality, artifacts, signature, watermark, username, blurry, more than one bridge, bad architecture"images = pipeline(prompt=prompt,negative_prompt=negative_prompt,width=512,height=512,num_inference_steps=25,num_images_per_prompt=4,generator=torch.manual_seed(0),
).imagesfrom diffusers.utils import make_image_gridmake_image_grid(images, 2, 2)

3.scheduler

scheduler没有参数化或训练；由配置文件定义。加载scheduler不会消耗大的内存，并且相同的配置文件可以用于各种不同的scheduler，比如下面的scheduler均可与StableDiffusionPipline兼容。

Diffusion流程本质上是由扩散模型和scheduler组成的集合，它们在一定程度上彼此独立。这意味着可以替换流程的某些部分，其中最好的例子就是scheduler。扩散模型通常只定义从噪声到较少噪声样本的前向传递过程，而调度器定义了整个去噪过程，包括：

去噪步骤是多少？随机的还是确定性的？用什么算法找到去噪样本？调度器可以非常复杂，并且经常在去噪速度和去噪质量之间进行权衡。

from diffusers import StableDiffusionPipeline
from diffusers import (DDPMScheduler,DDIMScheduler,PNDMScheduler,LMSDiscreteScheduler,EulerDiscreteScheduler,EulerAncestralDiscreteScheduler,DPMSolverMultistepScheduler,
)repo_id = "runwayml/stable-diffusion-v1-5"ddpm = DDPMScheduler.from_pretrained(repo_id, subfolder="scheduler")
ddim = DDIMScheduler.from_pretrained(repo_id, subfolder="scheduler")
pndm = PNDMScheduler.from_pretrained(repo_id, subfolder="scheduler")
lms = LMSDiscreteScheduler.from_pretrained(repo_id, subfolder="scheduler")
euler_anc = EulerAncestralDiscreteScheduler.from_pretrained(repo_id, subfolder="scheduler")
euler = EulerDiscreteScheduler.from_pretrained(repo_id, subfolder="scheduler")
dpm = DPMSolverMultistepScheduler.from_pretrained(repo_id, subfolder="scheduler")# replace `dpm` with any of `ddpm`, `ddim`, `pndm`, `lms`, `euler_anc`, `euler`
pipeline = StableDiffusionPipeline.from_pretrained(repo_id, scheduler=dpm, use_safetensors=True)

4.DiffusionPipline explained

作为一个类方法，DiffusionPipeline.from_pretrained()做两件事，1.下载推理所需的权重并缓存，一般存在在.cache文件中，2.将缓存文件中的model_index.json进行实例化。

feature_extractor--CLIPFeatureExtractor(transformers)；scheduler--PNDMScheduler；text_encoder--CLIPTextModel(transformers)；tokenizer--CLIPTokenizer(transformers)；unet--UNet2DConditionModel；vae--AutoencoderKL

{"_class_name": "StableDiffusionPipeline","_diffusers_version": "0.6.0","feature_extractor": ["transformers","CLIPImageProcessor"],"safety_checker": ["stable_diffusion","StableDiffusionSafetyChecker"],"scheduler": ["diffusers","PNDMScheduler"],"text_encoder": ["transformers","CLIPTextModel"],"tokenizer": ["transformers","CLIPTokenizer"],"unet": ["diffusers","UNet2DConditionModel"],"vae": ["diffusers","AutoencoderKL"]
}

下面是runway/stable-diffusion-v1-5的文件夹结构：

.
├── feature_extractor
│   └── preprocessor_config.json
├── model_index.json
├── safety_checker
│   ├── config.json
│   └── pytorch_model.bin
├── scheduler
│   └── scheduler_config.json
├── text_encoder
│   ├── config.json
│   └── pytorch_model.bin
├── tokenizer
│   ├── merges.txt
│   ├── special_tokens_map.json
│   ├── tokenizer_config.json
│   └── vocab.json
├── unet
│   ├── config.json
│   ├── diffusion_pytorch_model.bin
└── vae├── config.json├── diffusion_pytorch_model.bin

可以查看组件的属性和配置：

pipeline.tokenizer
CLIPTokenizer(name_or_path="/root/.cache/huggingface/hub/models--runwayml--stable-diffusion-v1-5/snapshots/39593d5650112b4cc580433f6b0435385882d819/tokenizer",vocab_size=49408,model_max_length=77,is_fast=False,padding_side="right",truncation_side="right",special_tokens={"bos_token": AddedToken("<|startoftext|>", rstrip=False, lstrip=False, single_word=False, normalized=True),"eos_token": AddedToken("<|endoftext|>", rstrip=False, lstrip=False, single_word=False, normalized=True),"unk_token": AddedToken("<|endoftext|>", rstrip=False, lstrip=False, single_word=False, normalized=True),"pad_token": "<|endoftext|>",},
)