README.md
| 1 | --- |
| 2 | tags: |
| 3 | - text-to-image |
| 4 | - stable-diffusion |
| 5 | |
| 6 | language: |
| 7 | - en |
| 8 | library_name: diffusers |
| 9 | --- |
| 10 | |
| 11 | # IP-Adapter-FaceID Model Card |
| 12 | |
| 13 | |
| 14 | <div align="center"> |
| 15 | |
| 16 | [**Project Page**](https://ip-adapter.github.io) **|** [**Paper (ArXiv)**](https://arxiv.org/abs/2308.06721) **|** [**Code**](https://github.com/tencent-ailab/IP-Adapter) |
| 17 | </div> |
| 18 | |
| 19 | --- |
| 20 | |
| 21 | |
| 22 | |
| 23 | ## Introduction |
| 24 | |
| 25 | An experimental version of IP-Adapter-FaceID: we use face ID embedding from a face recognition model instead of CLIP image embedding, additionally, we use LoRA to improve ID consistency. IP-Adapter-FaceID can generate various style images conditioned on a face with only text prompts. |
| 26 | |
| 27 |  |
| 28 | |
| 29 | |
| 30 | **Update 2023/12/27**: |
| 31 | |
| 32 | IP-Adapter-FaceID-Plus: face ID embedding (for face ID) + CLIP image embedding (for face structure) |
| 33 | |
| 34 | <div align="center"> |
| 35 | |
| 36 |  |
| 37 | </div> |
| 38 | |
| 39 | **Update 2023/12/28**: |
| 40 | |
| 41 | IP-Adapter-FaceID-PlusV2: face ID embedding (for face ID) + controllable CLIP image embedding (for face structure) |
| 42 | |
| 43 | You can adjust the weight of the face structure to get different generation! |
| 44 | |
| 45 | <div align="center"> |
| 46 | |
| 47 |  |
| 48 | </div> |
| 49 | |
| 50 | **Update 2024/01/04**: |
| 51 | |
| 52 | IP-Adapter-FaceID-SDXL: An experimental SDXL version of IP-Adapter-FaceID |
| 53 | |
| 54 | <div align="center"> |
| 55 | |
| 56 |  |
| 57 | </div> |
| 58 | |
| 59 | **Update 2024/01/17**: |
| 60 | |
| 61 | IP-Adapter-FaceID-PlusV2-SDXL: An experimental SDXL version of IP-Adapter-FaceID-PlusV2 |
| 62 | |
| 63 | |
| 64 | **Update 2024/01/19**: |
| 65 | |
| 66 | IP-Adapter-FaceID-Portrait: same with IP-Adapter-FaceID but for portrait generation (no lora! no controlnet!). Specifically, it accepts multiple facial images to enhance similarity (the default is 5). |
| 67 | |
| 68 | <div align="center"> |
| 69 | |
| 70 |  |
| 71 | </div> |
| 72 | |
| 73 | |
| 74 | ## Usage |
| 75 | |
| 76 | ### IP-Adapter-FaceID |
| 77 | |
| 78 | Firstly, you should use [insightface](https://github.com/deepinsight/insightface) to extract face ID embedding: |
| 79 | |
| 80 | ```python |
| 81 | |
| 82 | import cv2 |
| 83 | from insightface.app import FaceAnalysis |
| 84 | import torch |
| 85 | |
| 86 | app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider', 'CPUExecutionProvider']) |
| 87 | app.prepare(ctx_id=0, det_size=(640, 640)) |
| 88 | |
| 89 | image = cv2.imread("person.jpg") |
| 90 | faces = app.get(image) |
| 91 | |
| 92 | faceid_embeds = torch.from_numpy(faces[0].normed_embedding).unsqueeze(0) |
| 93 | ``` |
| 94 | |
| 95 | Then, you can generate images conditioned on the face embeddings: |
| 96 | |
| 97 | ```python |
| 98 | |
| 99 | import torch |
| 100 | from diffusers import StableDiffusionPipeline, DDIMScheduler, AutoencoderKL |
| 101 | from PIL import Image |
| 102 | |
| 103 | from ip_adapter.ip_adapter_faceid import IPAdapterFaceID |
| 104 | |
| 105 | base_model_path = "SG161222/Realistic_Vision_V4.0_noVAE" |
| 106 | vae_model_path = "stabilityai/sd-vae-ft-mse" |
| 107 | ip_ckpt = "ip-adapter-faceid_sd15.bin" |
| 108 | device = "cuda" |
| 109 | |
| 110 | noise_scheduler = DDIMScheduler( |
| 111 | num_train_timesteps=1000, |
| 112 | beta_start=0.00085, |
| 113 | beta_end=0.012, |
| 114 | beta_schedule="scaled_linear", |
| 115 | clip_sample=False, |
| 116 | set_alpha_to_one=False, |
| 117 | steps_offset=1, |
| 118 | ) |
| 119 | vae = AutoencoderKL.from_pretrained(vae_model_path).to(dtype=torch.float16) |
| 120 | pipe = StableDiffusionPipeline.from_pretrained( |
| 121 | base_model_path, |
| 122 | torch_dtype=torch.float16, |
| 123 | scheduler=noise_scheduler, |
| 124 | vae=vae, |
| 125 | feature_extractor=None, |
| 126 | safety_checker=None |
| 127 | ) |
| 128 | |
| 129 | # load ip-adapter |
| 130 | ip_model = IPAdapterFaceID(pipe, ip_ckpt, device) |
| 131 | |
| 132 | # generate image |
| 133 | prompt = "photo of a woman in red dress in a garden" |
| 134 | negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality, blurry" |
| 135 | |
| 136 | images = ip_model.generate( |
| 137 | prompt=prompt, negative_prompt=negative_prompt, faceid_embeds=faceid_embeds, num_samples=4, width=512, height=768, num_inference_steps=30, seed=2023 |
| 138 | ) |
| 139 | |
| 140 | ``` |
| 141 | |
| 142 | you can also use a normal IP-Adapter and a normal LoRA to load model: |
| 143 | |
| 144 | ```python |
| 145 | import torch |
| 146 | from diffusers import StableDiffusionPipeline, DDIMScheduler, AutoencoderKL |
| 147 | from PIL import Image |
| 148 | |
| 149 | from ip_adapter.ip_adapter_faceid_separate import IPAdapterFaceID |
| 150 | |
| 151 | base_model_path = "SG161222/Realistic_Vision_V4.0_noVAE" |
| 152 | vae_model_path = "stabilityai/sd-vae-ft-mse" |
| 153 | ip_ckpt = "ip-adapter-faceid_sd15.bin" |
| 154 | lora_ckpt = "ip-adapter-faceid_sd15_lora.safetensors" |
| 155 | device = "cuda" |
| 156 | |
| 157 | noise_scheduler = DDIMScheduler( |
| 158 | num_train_timesteps=1000, |
| 159 | beta_start=0.00085, |
| 160 | beta_end=0.012, |
| 161 | beta_schedule="scaled_linear", |
| 162 | clip_sample=False, |
| 163 | set_alpha_to_one=False, |
| 164 | steps_offset=1, |
| 165 | ) |
| 166 | vae = AutoencoderKL.from_pretrained(vae_model_path).to(dtype=torch.float16) |
| 167 | pipe = StableDiffusionPipeline.from_pretrained( |
| 168 | base_model_path, |
| 169 | torch_dtype=torch.float16, |
| 170 | scheduler=noise_scheduler, |
| 171 | vae=vae, |
| 172 | feature_extractor=None, |
| 173 | safety_checker=None |
| 174 | ) |
| 175 | |
| 176 | # load lora and fuse |
| 177 | pipe.load_lora_weights(lora_ckpt) |
| 178 | pipe.fuse_lora() |
| 179 | |
| 180 | # load ip-adapter |
| 181 | ip_model = IPAdapterFaceID(pipe, ip_ckpt, device) |
| 182 | |
| 183 | # generate image |
| 184 | prompt = "photo of a woman in red dress in a garden" |
| 185 | negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality, blurry" |
| 186 | |
| 187 | images = ip_model.generate( |
| 188 | prompt=prompt, negative_prompt=negative_prompt, faceid_embeds=faceid_embeds, num_samples=4, width=512, height=768, num_inference_steps=30, seed=2023 |
| 189 | ) |
| 190 | |
| 191 | |
| 192 | ``` |
| 193 | |
| 194 | ### IP-Adapter-FaceID-SDXL |
| 195 | |
| 196 | Firstly, you should use [insightface](https://github.com/deepinsight/insightface) to extract face ID embedding: |
| 197 | |
| 198 | ```python |
| 199 | |
| 200 | import cv2 |
| 201 | from insightface.app import FaceAnalysis |
| 202 | import torch |
| 203 | |
| 204 | app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider', 'CPUExecutionProvider']) |
| 205 | app.prepare(ctx_id=0, det_size=(640, 640)) |
| 206 | |
| 207 | image = cv2.imread("person.jpg") |
| 208 | faces = app.get(image) |
| 209 | |
| 210 | faceid_embeds = torch.from_numpy(faces[0].normed_embedding).unsqueeze(0) |
| 211 | ``` |
| 212 | |
| 213 | Then, you can generate images conditioned on the face embeddings: |
| 214 | |
| 215 | ```python |
| 216 | |
| 217 | import torch |
| 218 | from diffusers import StableDiffusionXLPipeline, DDIMScheduler |
| 219 | from PIL import Image |
| 220 | |
| 221 | from ip_adapter.ip_adapter_faceid import IPAdapterFaceIDXL |
| 222 | |
| 223 | base_model_path = "SG161222/RealVisXL_V3.0" |
| 224 | ip_ckpt = "ip-adapter-faceid_sdxl.bin" |
| 225 | device = "cuda" |
| 226 | |
| 227 | noise_scheduler = DDIMScheduler( |
| 228 | num_train_timesteps=1000, |
| 229 | beta_start=0.00085, |
| 230 | beta_end=0.012, |
| 231 | beta_schedule="scaled_linear", |
| 232 | clip_sample=False, |
| 233 | set_alpha_to_one=False, |
| 234 | steps_offset=1, |
| 235 | ) |
| 236 | pipe = StableDiffusionXLPipeline.from_pretrained( |
| 237 | base_model_path, |
| 238 | torch_dtype=torch.float16, |
| 239 | scheduler=noise_scheduler, |
| 240 | add_watermarker=False, |
| 241 | ) |
| 242 | |
| 243 | # load ip-adapter |
| 244 | ip_model = IPAdapterFaceIDXL(pipe, ip_ckpt, device) |
| 245 | |
| 246 | # generate image |
| 247 | 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" |
| 248 | negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality, blurry" |
| 249 | |
| 250 | images = ip_model.generate( |
| 251 | prompt=prompt, negative_prompt=negative_prompt, faceid_embeds=faceid_embeds, num_samples=2, |
| 252 | width=1024, height=1024, |
| 253 | num_inference_steps=30, guidance_scale=7.5, seed=2023 |
| 254 | ) |
| 255 | |
| 256 | ``` |
| 257 | |
| 258 | |
| 259 | ### IP-Adapter-FaceID-Plus |
| 260 | |
| 261 | Firstly, you should use [insightface](https://github.com/deepinsight/insightface) to extract face ID embedding and face image: |
| 262 | |
| 263 | ```python |
| 264 | |
| 265 | import cv2 |
| 266 | from insightface.app import FaceAnalysis |
| 267 | from insightface.utils import face_align |
| 268 | import torch |
| 269 | |
| 270 | app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider', 'CPUExecutionProvider']) |
| 271 | app.prepare(ctx_id=0, det_size=(640, 640)) |
| 272 | |
| 273 | image = cv2.imread("person.jpg") |
| 274 | faces = app.get(image) |
| 275 | |
| 276 | faceid_embeds = torch.from_numpy(faces[0].normed_embedding).unsqueeze(0) |
| 277 | face_image = face_align.norm_crop(image, landmark=faces[0].kps, image_size=224) # you can also segment the face |
| 278 | ``` |
| 279 | |
| 280 | Then, you can generate images conditioned on the face embeddings: |
| 281 | |
| 282 | ```python |
| 283 | |
| 284 | import torch |
| 285 | from diffusers import StableDiffusionPipeline, DDIMScheduler, AutoencoderKL |
| 286 | from PIL import Image |
| 287 | |
| 288 | from ip_adapter.ip_adapter_faceid import IPAdapterFaceIDPlus |
| 289 | |
| 290 | v2 = False |
| 291 | base_model_path = "SG161222/Realistic_Vision_V4.0_noVAE" |
| 292 | vae_model_path = "stabilityai/sd-vae-ft-mse" |
| 293 | image_encoder_path = "laion/CLIP-ViT-H-14-laion2B-s32B-b79K" |
| 294 | ip_ckpt = "ip-adapter-faceid-plus_sd15.bin" if not v2 else "ip-adapter-faceid-plusv2_sd15.bin" |
| 295 | device = "cuda" |
| 296 | |
| 297 | noise_scheduler = DDIMScheduler( |
| 298 | num_train_timesteps=1000, |
| 299 | beta_start=0.00085, |
| 300 | beta_end=0.012, |
| 301 | beta_schedule="scaled_linear", |
| 302 | clip_sample=False, |
| 303 | set_alpha_to_one=False, |
| 304 | steps_offset=1, |
| 305 | ) |
| 306 | vae = AutoencoderKL.from_pretrained(vae_model_path).to(dtype=torch.float16) |
| 307 | pipe = StableDiffusionPipeline.from_pretrained( |
| 308 | base_model_path, |
| 309 | torch_dtype=torch.float16, |
| 310 | scheduler=noise_scheduler, |
| 311 | vae=vae, |
| 312 | feature_extractor=None, |
| 313 | safety_checker=None |
| 314 | ) |
| 315 | |
| 316 | # load ip-adapter |
| 317 | ip_model = IPAdapterFaceIDPlus(pipe, image_encoder_path, ip_ckpt, device) |
| 318 | |
| 319 | # generate image |
| 320 | prompt = "photo of a woman in red dress in a garden" |
| 321 | negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality, blurry" |
| 322 | |
| 323 | images = ip_model.generate( |
| 324 | prompt=prompt, negative_prompt=negative_prompt, face_image=face_image, faceid_embeds=faceid_embeds, shortcut=v2, s_scale=1.0, |
| 325 | num_samples=4, width=512, height=768, num_inference_steps=30, seed=2023 |
| 326 | ) |
| 327 | |
| 328 | ``` |
| 329 | |
| 330 | ### IP-Adapter-FaceID-Portrait |
| 331 | |
| 332 | ```python |
| 333 | |
| 334 | import cv2 |
| 335 | from insightface.app import FaceAnalysis |
| 336 | import torch |
| 337 | |
| 338 | app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider', 'CPUExecutionProvider']) |
| 339 | app.prepare(ctx_id=0, det_size=(640, 640)) |
| 340 | |
| 341 | |
| 342 | images = ["1.jpg", "2.jpg", "3.jpg", "4.jpg", "5.jpg"] |
| 343 | |
| 344 | faceid_embeds = [] |
| 345 | for image in images: |
| 346 | image = cv2.imread("person.jpg") |
| 347 | faces = app.get(image) |
| 348 | faceid_embeds.append(torch.from_numpy(faces[0].normed_embedding).unsqueeze(0).unsqueeze(0)) |
| 349 | faceid_embeds = torch.cat(faceid_embeds, dim=1) |
| 350 | ``` |
| 351 | |
| 352 | ```python |
| 353 | import torch |
| 354 | from diffusers import StableDiffusionPipeline, DDIMScheduler, AutoencoderKL |
| 355 | from PIL import Image |
| 356 | |
| 357 | from ip_adapter.ip_adapter_faceid_separate import IPAdapterFaceID |
| 358 | |
| 359 | base_model_path = "SG161222/Realistic_Vision_V4.0_noVAE" |
| 360 | vae_model_path = "stabilityai/sd-vae-ft-mse" |
| 361 | ip_ckpt = "ip-adapter-faceid-portrait_sd15.bin" |
| 362 | device = "cuda" |
| 363 | |
| 364 | noise_scheduler = DDIMScheduler( |
| 365 | num_train_timesteps=1000, |
| 366 | beta_start=0.00085, |
| 367 | beta_end=0.012, |
| 368 | beta_schedule="scaled_linear", |
| 369 | clip_sample=False, |
| 370 | set_alpha_to_one=False, |
| 371 | steps_offset=1, |
| 372 | ) |
| 373 | vae = AutoencoderKL.from_pretrained(vae_model_path).to(dtype=torch.float16) |
| 374 | pipe = StableDiffusionPipeline.from_pretrained( |
| 375 | base_model_path, |
| 376 | torch_dtype=torch.float16, |
| 377 | scheduler=noise_scheduler, |
| 378 | vae=vae, |
| 379 | feature_extractor=None, |
| 380 | safety_checker=None |
| 381 | ) |
| 382 | |
| 383 | |
| 384 | # load ip-adapter |
| 385 | ip_model = IPAdapterFaceID(pipe, ip_ckpt, device, num_tokens=16, n_cond=5) |
| 386 | |
| 387 | # generate image |
| 388 | prompt = "photo of a woman in red dress in a garden" |
| 389 | negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality, blurry" |
| 390 | |
| 391 | images = ip_model.generate( |
| 392 | prompt=prompt, negative_prompt=negative_prompt, faceid_embeds=faceid_embeds, num_samples=4, width=512, height=512, num_inference_steps=30, seed=2023 |
| 393 | ) |
| 394 | |
| 395 | |
| 396 | ``` |
| 397 | |
| 398 | |
| 399 | |
| 400 | ## Limitations and Bias |
| 401 | - The models do not achieve perfect photorealism and ID consistency. |
| 402 | - The generalization of the models is limited due to limitations of the training data, base model and face recognition model. |
| 403 | |
| 404 | |
| 405 | ## Non-commercial use |
| 406 | **AS InsightFace pretrained models are available for non-commercial research purposes, IP-Adapter-FaceID models are released exclusively for research purposes and is not intended for commercial use.** |
| 407 | |
| 408 | |