IP-Adapter-FaceID 模型卡

项目页面 | 论文 (ArXiv) | 代码

简介

IP-Adapter-FaceID 的实验版：我们使用人脸识别模型的人脸 ID 嵌入，而不是 CLIP 图像嵌入，此外，我们使用 LoRA 来提高 ID 一致性。IP-Adapter-FaceID 可以仅使用文本提示生成各种风格的图像，并以人脸为条件。

results

2023 年 12 月 27 日更新：

IP-Adapter-FaceID-Plus：人脸 ID 嵌入（用于人脸 ID） + CLIP 图像嵌入（用于面部结构）

results

2023 年 12 月 28 日更新：

IP-Adapter-FaceID-PlusV2：人脸 ID 嵌入（用于人脸 ID） + 可控 CLIP 图像嵌入（用于面部结构）

您可以调整面部结构的权重以获得不同的生成结果！

results

2024 年 1 月 4 日更新：

IP-Adapter-FaceID-SDXL：IP-Adapter-FaceID 的实验性 SDXL 版本

results

2024 年 1 月 17 日更新：

IP-Adapter-FaceID-PlusV2-SDXL：IP-Adapter-FaceID-PlusV2 的实验性 SDXL 版本

2024 年 1 月 19 日更新：

IP-Adapter-FaceID-Portrait：与 IP-Adapter-FaceID 相同，但用于人像生成（没有 lora！没有 ControlNet！）。具体来说，它接受多张面部图像以增强相似性（默认值为 5）。

results

使用方法

IP-Adapter-FaceID

首先，您应该使用 insightface 提取人脸 ID 嵌入：


import cv2
from insightface.app import FaceAnalysis
import torch

app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
app.prepare(ctx_id=0, det_size=(640, 640))

image = cv2.imread("person.jpg")
faces = app.get(image)

faceid_embeds = torch.from_numpy(faces[0].normed_embedding).unsqueeze(0)

然后，您可以根据面部嵌入生成图像：


import torch
from diffusers import StableDiffusionPipeline, DDIMScheduler, AutoencoderKL
from PIL import Image

from ip_adapter.ip_adapter_faceid import IPAdapterFaceID

base_model_path = "SG161222/Realistic_Vision_V4.0_noVAE"
vae_model_path = "stabilityai/sd-vae-ft-mse"
ip_ckpt = "ip-adapter-faceid_sd15.bin"
device = "cuda"

noise_scheduler = DDIMScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    clip_sample=False,
    set_alpha_to_one=False,
    steps_offset=1,
)
vae = AutoencoderKL.from_pretrained(vae_model_path).to(dtype=torch.float16)
pipe = StableDiffusionPipeline.from_pretrained(
    base_model_path,
    torch_dtype=torch.float16,
    scheduler=noise_scheduler,
    vae=vae,
    feature_extractor=None,
    safety_checker=None
)

# load ip-adapter
ip_model = IPAdapterFaceID(pipe, ip_ckpt, device)

# generate image
prompt = "photo of a woman in red dress in a garden"
negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality, blurry"

images = ip_model.generate(
    prompt=prompt, negative_prompt=negative_prompt, faceid_embeds=faceid_embeds, num_samples=4, width=512, height=768, num_inference_steps=30, seed=2023
)

您也可以使用普通 IP-Adapter 和普通 LoRA 来加载模型：

import torch
from diffusers import StableDiffusionPipeline, DDIMScheduler, AutoencoderKL
from PIL import Image

from ip_adapter.ip_adapter_faceid_separate import IPAdapterFaceID

base_model_path = "SG161222/Realistic_Vision_V4.0_noVAE"
vae_model_path = "stabilityai/sd-vae-ft-mse"
ip_ckpt = "ip-adapter-faceid_sd15.bin"
lora_ckpt = "ip-adapter-faceid_sd15_lora.safetensors"
device = "cuda"

noise_scheduler = DDIMScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    clip_sample=False,
    set_alpha_to_one=False,
    steps_offset=1,
)
vae = AutoencoderKL.from_pretrained(vae_model_path).to(dtype=torch.float16)
pipe = StableDiffusionPipeline.from_pretrained(
    base_model_path,
    torch_dtype=torch.float16,
    scheduler=noise_scheduler,
    vae=vae,
    feature_extractor=None,
    safety_checker=None
)

# load lora and fuse
pipe.load_lora_weights(lora_ckpt)
pipe.fuse_lora()

# load ip-adapter
ip_model = IPAdapterFaceID(pipe, ip_ckpt, device)

# generate image
prompt = "photo of a woman in red dress in a garden"
negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality, blurry"

images = ip_model.generate(
    prompt=prompt, negative_prompt=negative_prompt, faceid_embeds=faceid_embeds, num_samples=4, width=512, height=768, num_inference_steps=30, seed=2023
)

IP-Adapter-FaceID-SDXL

首先，您应该使用 insightface 提取 Face ID 向量：


import cv2
from insightface.app import FaceAnalysis
import torch

app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
app.prepare(ctx_id=0, det_size=(640, 640))

image = cv2.imread("person.jpg")
faces = app.get(image)

faceid_embeds = torch.from_numpy(faces[0].normed_embedding).unsqueeze(0)

然后，您可以根据人脸嵌入生成图像：


import torch
from diffusers import StableDiffusionXLPipeline, DDIMScheduler
from PIL import Image

from ip_adapter.ip_adapter_faceid import IPAdapterFaceIDXL

base_model_path = "SG161222/RealVisXL_V3.0"
ip_ckpt = "ip-adapter-faceid_sdxl.bin"
device = "cuda"

noise_scheduler = DDIMScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    clip_sample=False,
    set_alpha_to_one=False,
    steps_offset=1,
)
pipe = StableDiffusionXLPipeline.from_pretrained(
    base_model_path,
    torch_dtype=torch.float16,
    scheduler=noise_scheduler,
    add_watermarker=False,
)

# load ip-adapter
ip_model = IPAdapterFaceIDXL(pipe, ip_ckpt, device)

# generate image
prompt = "A closeup shot of a beautiful Asian teenage girl in a white dress wearing small silver earrings in the garden, under the soft morning light"
negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality, blurry"

images = ip_model.generate(
    prompt=prompt, negative_prompt=negative_prompt, faceid_embeds=faceid_embeds, num_samples=2,
    width=1024, height=1024,
    num_inference_steps=30, guidance_scale=7.5, seed=2023
)

IP-Adapter-FaceID-Plus

首先，您应该使用 insightface 提取人脸 ID 向量和人脸图像：


import cv2
from insightface.app import FaceAnalysis
from insightface.utils import face_align
import torch

app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
app.prepare(ctx_id=0, det_size=(640, 640))

image = cv2.imread("person.jpg")
faces = app.get(image)

faceid_embeds = torch.from_numpy(faces[0].normed_embedding).unsqueeze(0)
face_image = face_align.norm_crop(image, landmark=faces[0].kps, image_size=224) # you can also segment the face

然后，您可以根据人脸嵌入生成图像：


import torch
from diffusers import StableDiffusionPipeline, DDIMScheduler, AutoencoderKL
from PIL import Image

from ip_adapter.ip_adapter_faceid import IPAdapterFaceIDPlus

v2 = False
base_model_path = "SG161222/Realistic_Vision_V4.0_noVAE"
vae_model_path = "stabilityai/sd-vae-ft-mse"
image_encoder_path = "laion/CLIP-ViT-H-14-laion2B-s32B-b79K"
ip_ckpt = "ip-adapter-faceid-plus_sd15.bin" if not v2 else "ip-adapter-faceid-plusv2_sd15.bin"
device = "cuda"

noise_scheduler = DDIMScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    clip_sample=False,
    set_alpha_to_one=False,
    steps_offset=1,
)
vae = AutoencoderKL.from_pretrained(vae_model_path).to(dtype=torch.float16)
pipe = StableDiffusionPipeline.from_pretrained(
    base_model_path,
    torch_dtype=torch.float16,
    scheduler=noise_scheduler,
    vae=vae,
    feature_extractor=None,
    safety_checker=None
)

# load ip-adapter
ip_model = IPAdapterFaceIDPlus(pipe, image_encoder_path, ip_ckpt, device)

# generate image
prompt = "photo of a woman in red dress in a garden"
negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality, blurry"

images = ip_model.generate(
     prompt=prompt, negative_prompt=negative_prompt, face_image=face_image, faceid_embeds=faceid_embeds, shortcut=v2, s_scale=1.0,
     num_samples=4, width=512, height=768, num_inference_steps=30, seed=2023
)

IP 适配器 - FaceID 人脸识别


import cv2
from insightface.app import FaceAnalysis
import torch

app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
app.prepare(ctx_id=0, det_size=(640, 640))


images = ["1.jpg", "2.jpg", "3.jpg", "4.jpg", "5.jpg"]

faceid_embeds = []
for image in images:
    image = cv2.imread("person.jpg")
    faces = app.get(image)
    faceid_embeds.append(torch.from_numpy(faces[0].normed_embedding).unsqueeze(0).unsqueeze(0))
  faceid_embeds = torch.cat(faceid_embeds, dim=1)

好的，请提供您要翻译的文本。我会尽力将它翻译成通俗、专业、优雅且流畅的中文，并保持原始的 Markdown 格式。

import torch
from diffusers import StableDiffusionPipeline, DDIMScheduler, AutoencoderKL
from PIL import Image

from ip_adapter.ip_adapter_faceid_separate import IPAdapterFaceID

base_model_path = "SG161222/Realistic_Vision_V4.0_noVAE"
vae_model_path = "stabilityai/sd-vae-ft-mse"
ip_ckpt = "ip-adapter-faceid-portrait_sd15.bin"
device = "cuda"

noise_scheduler = DDIMScheduler(
    num_train_timesteps=1000,
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    clip_sample=False,
    set_alpha_to_one=False,
    steps_offset=1,
)
vae = AutoencoderKL.from_pretrained(vae_model_path).to(dtype=torch.float16)
pipe = StableDiffusionPipeline.from_pretrained(
    base_model_path,
    torch_dtype=torch.float16,
    scheduler=noise_scheduler,
    vae=vae,
    feature_extractor=None,
    safety_checker=None
)


# load ip-adapter
ip_model = IPAdapterFaceID(pipe, ip_ckpt, device, num_tokens=16, n_cond=5)

# generate image
prompt = "photo of a woman in red dress in a garden"
negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality, blurry"

images = ip_model.generate(
    prompt=prompt, negative_prompt=negative_prompt, faceid_embeds=faceid_embeds, num_samples=4, width=512, height=512, num_inference_steps=30, seed=2023
)

限制和偏差

模型未能达到完美的逼真度和身份一致性。
由于训练数据、基础模型和人脸识别模型的局限性，模型的泛化能力有限。

非商业用途

AS InsightFace 预训练模型可用于非商业研究目的，IP-Adapter-FaceID 模型仅供研究目的使用，不适用于商业用途。