Spaces:
Runtime error
Runtime error
rodrigomasini
commited on
Commit
•
96f7484
1
Parent(s):
093740e
Update app.py
Browse files
app.py
CHANGED
@@ -1001,7 +1001,579 @@ use_safetensors= False
|
|
1001 |
# # pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
|
1002 |
# pipe1.scheduler.set_timesteps(50)
|
1003 |
###
|
1004 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1005 |
models_dict["Unstable"], torch_dtype=torch.float16, use_safetensors=use_safetensors)
|
1006 |
pipe2 = pipe2.to("cpu")
|
1007 |
pipe2.load_photomaker_adapter(
|
|
|
1001 |
# # pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
|
1002 |
# pipe1.scheduler.set_timesteps(50)
|
1003 |
###
|
1004 |
+
|
1005 |
+
from typing import Any, Callable, Dict, List, Optional, Union, Tuple
|
1006 |
+
from collections import OrderedDict
|
1007 |
+
import os
|
1008 |
+
import PIL
|
1009 |
+
import numpy as np
|
1010 |
+
|
1011 |
+
import torch
|
1012 |
+
from torchvision import transforms as T
|
1013 |
+
|
1014 |
+
from safetensors import safe_open
|
1015 |
+
from huggingface_hub.utils import validate_hf_hub_args
|
1016 |
+
from transformers import CLIPImageProcessor, CLIPTokenizer
|
1017 |
+
from diffusers import StableDiffusionXLPipeline
|
1018 |
+
from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
|
1019 |
+
from diffusers.utils import (
|
1020 |
+
_get_model_file,
|
1021 |
+
is_transformers_available,
|
1022 |
+
logging,
|
1023 |
+
)
|
1024 |
+
|
1025 |
+
from . import PhotoMakerIDEncoder
|
1026 |
+
|
1027 |
+
PipelineImageInput = Union[
|
1028 |
+
PIL.Image.Image,
|
1029 |
+
torch.FloatTensor,
|
1030 |
+
List[PIL.Image.Image],
|
1031 |
+
List[torch.FloatTensor],
|
1032 |
+
]
|
1033 |
+
|
1034 |
+
|
1035 |
+
class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
|
1036 |
+
@validate_hf_hub_args
|
1037 |
+
def load_photomaker_adapter(
|
1038 |
+
self,
|
1039 |
+
pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
|
1040 |
+
weight_name: str,
|
1041 |
+
subfolder: str = '',
|
1042 |
+
trigger_word: str = 'img',
|
1043 |
+
**kwargs,
|
1044 |
+
):
|
1045 |
+
"""
|
1046 |
+
Parameters:
|
1047 |
+
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
|
1048 |
+
Can be either:
|
1049 |
+
- A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
|
1050 |
+
the Hub.
|
1051 |
+
- A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
|
1052 |
+
with [`ModelMixin.save_pretrained`].
|
1053 |
+
- A [torch state
|
1054 |
+
dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
|
1055 |
+
weight_name (`str`):
|
1056 |
+
The weight name NOT the path to the weight.
|
1057 |
+
subfolder (`str`, defaults to `""`):
|
1058 |
+
The subfolder location of a model file within a larger model repository on the Hub or locally.
|
1059 |
+
trigger_word (`str`, *optional*, defaults to `"img"`):
|
1060 |
+
The trigger word is used to identify the position of class word in the text prompt,
|
1061 |
+
and it is recommended not to set it as a common word.
|
1062 |
+
This trigger word must be placed after the class word when used, otherwise, it will affect the performance of the personalized generation.
|
1063 |
+
"""
|
1064 |
+
|
1065 |
+
# Load the main state dict first.
|
1066 |
+
cache_dir = kwargs.pop("cache_dir", None)
|
1067 |
+
force_download = kwargs.pop("force_download", False)
|
1068 |
+
resume_download = kwargs.pop("resume_download", False)
|
1069 |
+
proxies = kwargs.pop("proxies", None)
|
1070 |
+
local_files_only = kwargs.pop("local_files_only", None)
|
1071 |
+
token = kwargs.pop("token", None)
|
1072 |
+
revision = kwargs.pop("revision", None)
|
1073 |
+
|
1074 |
+
user_agent = {
|
1075 |
+
"file_type": "attn_procs_weights",
|
1076 |
+
"framework": "pytorch",
|
1077 |
+
}
|
1078 |
+
|
1079 |
+
if not isinstance(pretrained_model_name_or_path_or_dict, dict):
|
1080 |
+
model_file = _get_model_file(
|
1081 |
+
pretrained_model_name_or_path_or_dict,
|
1082 |
+
weights_name=weight_name,
|
1083 |
+
cache_dir=cache_dir,
|
1084 |
+
force_download=force_download,
|
1085 |
+
resume_download=resume_download,
|
1086 |
+
proxies=proxies,
|
1087 |
+
local_files_only=local_files_only,
|
1088 |
+
token=token,
|
1089 |
+
revision=revision,
|
1090 |
+
subfolder=subfolder,
|
1091 |
+
user_agent=user_agent,
|
1092 |
+
)
|
1093 |
+
if weight_name.endswith(".safetensors"):
|
1094 |
+
state_dict = {"id_encoder": {}, "lora_weights": {}}
|
1095 |
+
with safe_open(model_file, framework="pt", device="cpu") as f:
|
1096 |
+
for key in f.keys():
|
1097 |
+
if key.startswith("id_encoder."):
|
1098 |
+
state_dict["id_encoder"][key.replace("id_encoder.", "")] = f.get_tensor(key)
|
1099 |
+
elif key.startswith("lora_weights."):
|
1100 |
+
state_dict["lora_weights"][key.replace("lora_weights.", "")] = f.get_tensor(key)
|
1101 |
+
else:
|
1102 |
+
state_dict = torch.load(model_file, map_location="cpu")
|
1103 |
+
else:
|
1104 |
+
state_dict = pretrained_model_name_or_path_or_dict
|
1105 |
+
|
1106 |
+
keys = list(state_dict.keys())
|
1107 |
+
if keys != ["id_encoder", "lora_weights"]:
|
1108 |
+
raise ValueError("Required keys are (`id_encoder` and `lora_weights`) missing from the state dict.")
|
1109 |
+
|
1110 |
+
self.trigger_word = trigger_word
|
1111 |
+
# load finetuned CLIP image encoder and fuse module here if it has not been registered to the pipeline yet
|
1112 |
+
print(f"Loading PhotoMaker components [1] id_encoder from [{pretrained_model_name_or_path_or_dict}]...")
|
1113 |
+
id_encoder = PhotoMakerIDEncoder()
|
1114 |
+
id_encoder.load_state_dict(state_dict["id_encoder"], strict=True)
|
1115 |
+
id_encoder = id_encoder.to(self.device, dtype=self.unet.dtype)
|
1116 |
+
self.id_encoder = id_encoder
|
1117 |
+
self.id_image_processor = CLIPImageProcessor()
|
1118 |
+
|
1119 |
+
# load lora into models
|
1120 |
+
print(f"Loading PhotoMaker components [2] lora_weights from [{pretrained_model_name_or_path_or_dict}]")
|
1121 |
+
self.load_lora_weights(state_dict["lora_weights"], adapter_name="photomaker")
|
1122 |
+
|
1123 |
+
# Add trigger word token
|
1124 |
+
if self.tokenizer is not None:
|
1125 |
+
self.tokenizer.add_tokens([self.trigger_word], special_tokens=True)
|
1126 |
+
|
1127 |
+
self.tokenizer_2.add_tokens([self.trigger_word], special_tokens=True)
|
1128 |
+
|
1129 |
+
|
1130 |
+
def encode_prompt_with_trigger_word(
|
1131 |
+
self,
|
1132 |
+
prompt: str,
|
1133 |
+
prompt_2: Optional[str] = None,
|
1134 |
+
num_id_images: int = 1,
|
1135 |
+
device: Optional[torch.device] = None,
|
1136 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
1137 |
+
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
1138 |
+
class_tokens_mask: Optional[torch.LongTensor] = None,
|
1139 |
+
):
|
1140 |
+
device = device or self._execution_device
|
1141 |
+
|
1142 |
+
if prompt is not None and isinstance(prompt, str):
|
1143 |
+
batch_size = 1
|
1144 |
+
elif prompt is not None and isinstance(prompt, list):
|
1145 |
+
batch_size = len(prompt)
|
1146 |
+
else:
|
1147 |
+
batch_size = prompt_embeds.shape[0]
|
1148 |
+
|
1149 |
+
# Find the token id of the trigger word
|
1150 |
+
image_token_id = self.tokenizer_2.convert_tokens_to_ids(self.trigger_word)
|
1151 |
+
|
1152 |
+
# Define tokenizers and text encoders
|
1153 |
+
tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
|
1154 |
+
text_encoders = (
|
1155 |
+
[self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
|
1156 |
+
)
|
1157 |
+
|
1158 |
+
if prompt_embeds is None:
|
1159 |
+
prompt_2 = prompt_2 or prompt
|
1160 |
+
prompt_embeds_list = []
|
1161 |
+
prompts = [prompt, prompt_2]
|
1162 |
+
for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
|
1163 |
+
input_ids = tokenizer.encode(prompt) # TODO: batch encode
|
1164 |
+
clean_index = 0
|
1165 |
+
clean_input_ids = []
|
1166 |
+
class_token_index = []
|
1167 |
+
# Find out the corresponding class word token based on the newly added trigger word token
|
1168 |
+
for i, token_id in enumerate(input_ids):
|
1169 |
+
if token_id == image_token_id:
|
1170 |
+
class_token_index.append(clean_index - 1)
|
1171 |
+
else:
|
1172 |
+
clean_input_ids.append(token_id)
|
1173 |
+
clean_index += 1
|
1174 |
+
|
1175 |
+
if len(class_token_index) != 1:
|
1176 |
+
raise ValueError(
|
1177 |
+
f"PhotoMaker currently does not support multiple trigger words in a single prompt.\
|
1178 |
+
Trigger word: {self.trigger_word}, Prompt: {prompt}."
|
1179 |
+
)
|
1180 |
+
class_token_index = class_token_index[0]
|
1181 |
+
|
1182 |
+
# Expand the class word token and corresponding mask
|
1183 |
+
class_token = clean_input_ids[class_token_index]
|
1184 |
+
clean_input_ids = clean_input_ids[:class_token_index] + [class_token] * num_id_images + \
|
1185 |
+
clean_input_ids[class_token_index+1:]
|
1186 |
+
|
1187 |
+
# Truncation or padding
|
1188 |
+
max_len = tokenizer.model_max_length
|
1189 |
+
if len(clean_input_ids) > max_len:
|
1190 |
+
clean_input_ids = clean_input_ids[:max_len]
|
1191 |
+
else:
|
1192 |
+
clean_input_ids = clean_input_ids + [tokenizer.pad_token_id] * (
|
1193 |
+
max_len - len(clean_input_ids)
|
1194 |
+
)
|
1195 |
+
|
1196 |
+
class_tokens_mask = [True if class_token_index <= i < class_token_index+num_id_images else False \
|
1197 |
+
for i in range(len(clean_input_ids))]
|
1198 |
+
|
1199 |
+
clean_input_ids = torch.tensor(clean_input_ids, dtype=torch.long).unsqueeze(0)
|
1200 |
+
class_tokens_mask = torch.tensor(class_tokens_mask, dtype=torch.bool).unsqueeze(0)
|
1201 |
+
|
1202 |
+
prompt_embeds = text_encoder(
|
1203 |
+
clean_input_ids.to(device),
|
1204 |
+
output_hidden_states=True,
|
1205 |
+
)
|
1206 |
+
|
1207 |
+
# We are only ALWAYS interested in the pooled output of the final text encoder
|
1208 |
+
pooled_prompt_embeds = prompt_embeds[0]
|
1209 |
+
prompt_embeds = prompt_embeds.hidden_states[-2]
|
1210 |
+
prompt_embeds_list.append(prompt_embeds)
|
1211 |
+
|
1212 |
+
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
|
1213 |
+
|
1214 |
+
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
|
1215 |
+
class_tokens_mask = class_tokens_mask.to(device=device) # TODO: ignoring two-prompt case
|
1216 |
+
|
1217 |
+
return prompt_embeds, pooled_prompt_embeds, class_tokens_mask
|
1218 |
+
|
1219 |
+
@property
|
1220 |
+
def interrupt(self):
|
1221 |
+
return self._interrupt
|
1222 |
+
|
1223 |
+
@torch.no_grad()
|
1224 |
+
def __call__(
|
1225 |
+
self,
|
1226 |
+
prompt: Union[str, List[str]] = None,
|
1227 |
+
prompt_2: Optional[Union[str, List[str]]] = None,
|
1228 |
+
height: Optional[int] = None,
|
1229 |
+
width: Optional[int] = None,
|
1230 |
+
num_inference_steps: int = 50,
|
1231 |
+
denoising_end: Optional[float] = None,
|
1232 |
+
guidance_scale: float = 5.0,
|
1233 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
1234 |
+
negative_prompt_2: Optional[Union[str, List[str]]] = None,
|
1235 |
+
num_images_per_prompt: Optional[int] = 1,
|
1236 |
+
eta: float = 0.0,
|
1237 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
1238 |
+
latents: Optional[torch.FloatTensor] = None,
|
1239 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
1240 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
1241 |
+
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
1242 |
+
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
1243 |
+
output_type: Optional[str] = "pil",
|
1244 |
+
return_dict: bool = True,
|
1245 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
1246 |
+
guidance_rescale: float = 0.0,
|
1247 |
+
original_size: Optional[Tuple[int, int]] = None,
|
1248 |
+
crops_coords_top_left: Tuple[int, int] = (0, 0),
|
1249 |
+
target_size: Optional[Tuple[int, int]] = None,
|
1250 |
+
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
1251 |
+
callback_steps: int = 1,
|
1252 |
+
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
1253 |
+
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
1254 |
+
# Added parameters (for PhotoMaker)
|
1255 |
+
input_id_images: PipelineImageInput = None,
|
1256 |
+
start_merge_step: int = 0, # TODO: change to `style_strength_ratio` in the future
|
1257 |
+
class_tokens_mask: Optional[torch.LongTensor] = None,
|
1258 |
+
prompt_embeds_text_only: Optional[torch.FloatTensor] = None,
|
1259 |
+
pooled_prompt_embeds_text_only: Optional[torch.FloatTensor] = None,
|
1260 |
+
):
|
1261 |
+
r"""
|
1262 |
+
Function invoked when calling the pipeline for generation.
|
1263 |
+
Only the parameters introduced by PhotoMaker are discussed here.
|
1264 |
+
For explanations of the previous parameters in StableDiffusionXLPipeline, please refer to https://github.com/huggingface/diffusers/blob/v0.25.0/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py
|
1265 |
+
Args:
|
1266 |
+
input_id_images (`PipelineImageInput`, *optional*):
|
1267 |
+
Input ID Image to work with PhotoMaker.
|
1268 |
+
class_tokens_mask (`torch.LongTensor`, *optional*):
|
1269 |
+
Pre-generated class token. When the `prompt_embeds` parameter is provided in advance, it is necessary to prepare the `class_tokens_mask` beforehand for marking out the position of class word.
|
1270 |
+
prompt_embeds_text_only (`torch.FloatTensor`, *optional*):
|
1271 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
1272 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
1273 |
+
pooled_prompt_embeds_text_only (`torch.FloatTensor`, *optional*):
|
1274 |
+
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
|
1275 |
+
If not provided, pooled text embeddings will be generated from `prompt` input argument.
|
1276 |
+
Returns:
|
1277 |
+
[`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
|
1278 |
+
[`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
|
1279 |
+
`tuple`. When returning a tuple, the first element is a list with the generated images.
|
1280 |
+
"""
|
1281 |
+
# 0. Default height and width to unet
|
1282 |
+
height = height or self.unet.config.sample_size * self.vae_scale_factor
|
1283 |
+
width = width or self.unet.config.sample_size * self.vae_scale_factor
|
1284 |
+
|
1285 |
+
original_size = original_size or (height, width)
|
1286 |
+
target_size = target_size or (height, width)
|
1287 |
+
|
1288 |
+
# 1. Check inputs. Raise error if not correct
|
1289 |
+
self.check_inputs(
|
1290 |
+
prompt,
|
1291 |
+
prompt_2,
|
1292 |
+
height,
|
1293 |
+
width,
|
1294 |
+
callback_steps,
|
1295 |
+
negative_prompt,
|
1296 |
+
negative_prompt_2,
|
1297 |
+
prompt_embeds,
|
1298 |
+
negative_prompt_embeds,
|
1299 |
+
pooled_prompt_embeds,
|
1300 |
+
negative_pooled_prompt_embeds,
|
1301 |
+
callback_on_step_end_tensor_inputs,
|
1302 |
+
)
|
1303 |
+
|
1304 |
+
self._interrupt = False
|
1305 |
+
|
1306 |
+
#
|
1307 |
+
if prompt_embeds is not None and class_tokens_mask is None:
|
1308 |
+
raise ValueError(
|
1309 |
+
"If `prompt_embeds` are provided, `class_tokens_mask` also have to be passed. Make sure to generate `class_tokens_mask` from the same tokenizer that was used to generate `prompt_embeds`."
|
1310 |
+
)
|
1311 |
+
# check the input id images
|
1312 |
+
if input_id_images is None:
|
1313 |
+
raise ValueError(
|
1314 |
+
"Provide `input_id_images`. Cannot leave `input_id_images` undefined for PhotoMaker pipeline."
|
1315 |
+
)
|
1316 |
+
if not isinstance(input_id_images, list):
|
1317 |
+
input_id_images = [input_id_images]
|
1318 |
+
|
1319 |
+
# 2. Define call parameters
|
1320 |
+
if prompt is not None and isinstance(prompt, str):
|
1321 |
+
batch_size = 1
|
1322 |
+
prompt = [prompt]
|
1323 |
+
elif prompt is not None and isinstance(prompt, list):
|
1324 |
+
batch_size = len(prompt)
|
1325 |
+
else:
|
1326 |
+
batch_size = prompt_embeds.shape[0]
|
1327 |
+
|
1328 |
+
device = self._execution_device
|
1329 |
+
|
1330 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
1331 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
1332 |
+
# corresponds to doing no classifier free guidance.
|
1333 |
+
do_classifier_free_guidance = guidance_scale >= 1.0
|
1334 |
+
|
1335 |
+
assert do_classifier_free_guidance
|
1336 |
+
|
1337 |
+
# 3. Encode input prompt
|
1338 |
+
num_id_images = len(input_id_images)
|
1339 |
+
if isinstance(prompt, list):
|
1340 |
+
prompt_arr = prompt
|
1341 |
+
negative_prompt_embeds_arr = []
|
1342 |
+
prompt_embeds_text_only_arr = []
|
1343 |
+
prompt_embeds_arr = []
|
1344 |
+
latents_arr = []
|
1345 |
+
add_time_ids_arr = []
|
1346 |
+
negative_pooled_prompt_embeds_arr = []
|
1347 |
+
pooled_prompt_embeds_text_only_arr = []
|
1348 |
+
pooled_prompt_embeds_arr = []
|
1349 |
+
for prompt in prompt_arr:
|
1350 |
+
(
|
1351 |
+
prompt_embeds,
|
1352 |
+
pooled_prompt_embeds,
|
1353 |
+
class_tokens_mask,
|
1354 |
+
) = self.encode_prompt_with_trigger_word(
|
1355 |
+
prompt=prompt,
|
1356 |
+
prompt_2=prompt_2,
|
1357 |
+
device=device,
|
1358 |
+
num_id_images=num_id_images,
|
1359 |
+
prompt_embeds=prompt_embeds,
|
1360 |
+
pooled_prompt_embeds=pooled_prompt_embeds,
|
1361 |
+
class_tokens_mask=class_tokens_mask,
|
1362 |
+
)
|
1363 |
+
|
1364 |
+
# 4. Encode input prompt without the trigger word for delayed conditioning
|
1365 |
+
# encode, remove trigger word token, then decode
|
1366 |
+
tokens_text_only = self.tokenizer.encode(prompt, add_special_tokens=False)
|
1367 |
+
trigger_word_token = self.tokenizer.convert_tokens_to_ids(self.trigger_word)
|
1368 |
+
tokens_text_only.remove(trigger_word_token)
|
1369 |
+
prompt_text_only = self.tokenizer.decode(tokens_text_only, add_special_tokens=False)
|
1370 |
+
print(prompt_text_only)
|
1371 |
+
(
|
1372 |
+
prompt_embeds_text_only,
|
1373 |
+
negative_prompt_embeds,
|
1374 |
+
pooled_prompt_embeds_text_only, # TODO: replace the pooled_prompt_embeds with text only prompt
|
1375 |
+
negative_pooled_prompt_embeds,
|
1376 |
+
) = self.encode_prompt(
|
1377 |
+
prompt=prompt_text_only,
|
1378 |
+
prompt_2=prompt_2,
|
1379 |
+
device=device,
|
1380 |
+
num_images_per_prompt=num_images_per_prompt,
|
1381 |
+
do_classifier_free_guidance=True,
|
1382 |
+
negative_prompt=negative_prompt,
|
1383 |
+
negative_prompt_2=negative_prompt_2,
|
1384 |
+
prompt_embeds=prompt_embeds_text_only,
|
1385 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
1386 |
+
pooled_prompt_embeds=pooled_prompt_embeds_text_only,
|
1387 |
+
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
1388 |
+
)
|
1389 |
+
|
1390 |
+
# 5. Prepare the input ID images
|
1391 |
+
dtype = next(self.id_encoder.parameters()).dtype
|
1392 |
+
if not isinstance(input_id_images[0], torch.Tensor):
|
1393 |
+
id_pixel_values = self.id_image_processor(input_id_images, return_tensors="pt").pixel_values
|
1394 |
+
|
1395 |
+
id_pixel_values = id_pixel_values.unsqueeze(0).to(device=device, dtype=dtype) # TODO: multiple prompts
|
1396 |
+
|
1397 |
+
# 6. Get the update text embedding with the stacked ID embedding
|
1398 |
+
prompt_embeds = self.id_encoder(id_pixel_values, prompt_embeds, class_tokens_mask)
|
1399 |
+
|
1400 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
1401 |
+
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
1402 |
+
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
1403 |
+
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
1404 |
+
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
|
1405 |
+
bs_embed * num_images_per_prompt, -1
|
1406 |
+
)
|
1407 |
+
|
1408 |
+
|
1409 |
+
negative_prompt_embeds_arr.append(negative_prompt_embeds)
|
1410 |
+
negative_prompt_embeds = None
|
1411 |
+
negative_pooled_prompt_embeds_arr.append(negative_pooled_prompt_embeds)
|
1412 |
+
negative_pooled_prompt_embeds = None
|
1413 |
+
prompt_embeds_text_only_arr.append(prompt_embeds_text_only)
|
1414 |
+
prompt_embeds_text_only = None
|
1415 |
+
prompt_embeds_arr.append(prompt_embeds)
|
1416 |
+
prompt_embeds = None
|
1417 |
+
pooled_prompt_embeds_arr.append(pooled_prompt_embeds)
|
1418 |
+
pooled_prompt_embeds = None
|
1419 |
+
pooled_prompt_embeds_text_only_arr.append(pooled_prompt_embeds_text_only)
|
1420 |
+
pooled_prompt_embeds_text_only = None
|
1421 |
+
# 7. Prepare timesteps
|
1422 |
+
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
1423 |
+
timesteps = self.scheduler.timesteps
|
1424 |
+
|
1425 |
+
negative_prompt_embeds = torch.cat(negative_prompt_embeds_arr ,dim =0)
|
1426 |
+
print(negative_prompt_embeds.shape)
|
1427 |
+
prompt_embeds = torch.cat(prompt_embeds_arr ,dim = 0)
|
1428 |
+
print(prompt_embeds.shape)
|
1429 |
+
|
1430 |
+
prompt_embeds_text_only = torch.cat(prompt_embeds_text_only_arr ,dim = 0)
|
1431 |
+
print(prompt_embeds_text_only.shape)
|
1432 |
+
pooled_prompt_embeds_text_only = torch.cat(pooled_prompt_embeds_text_only_arr ,dim = 0)
|
1433 |
+
print(pooled_prompt_embeds_text_only.shape)
|
1434 |
+
|
1435 |
+
negative_pooled_prompt_embeds = torch.cat(negative_pooled_prompt_embeds_arr ,dim = 0)
|
1436 |
+
print(negative_pooled_prompt_embeds.shape)
|
1437 |
+
pooled_prompt_embeds = torch.cat(pooled_prompt_embeds_arr,dim = 0)
|
1438 |
+
print(pooled_prompt_embeds.shape)
|
1439 |
+
# 8. Prepare latent variables
|
1440 |
+
num_channels_latents = self.unet.config.in_channels
|
1441 |
+
latents = self.prepare_latents(
|
1442 |
+
batch_size * num_images_per_prompt,
|
1443 |
+
num_channels_latents,
|
1444 |
+
height,
|
1445 |
+
width,
|
1446 |
+
prompt_embeds.dtype,
|
1447 |
+
device,
|
1448 |
+
generator,
|
1449 |
+
latents,
|
1450 |
+
)
|
1451 |
+
|
1452 |
+
# 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
1453 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
1454 |
+
|
1455 |
+
# 10. Prepare added time ids & embeddings
|
1456 |
+
if self.text_encoder_2 is None:
|
1457 |
+
text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
|
1458 |
+
else:
|
1459 |
+
text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
|
1460 |
+
|
1461 |
+
add_time_ids = self._get_add_time_ids(
|
1462 |
+
original_size,
|
1463 |
+
crops_coords_top_left,
|
1464 |
+
target_size,
|
1465 |
+
dtype=prompt_embeds.dtype,
|
1466 |
+
text_encoder_projection_dim=text_encoder_projection_dim,
|
1467 |
+
)
|
1468 |
+
add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
|
1469 |
+
add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
|
1470 |
+
|
1471 |
+
|
1472 |
+
print(latents.shape)
|
1473 |
+
print(add_time_ids.shape)
|
1474 |
+
|
1475 |
+
# 11. Denoising loop
|
1476 |
+
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
|
1477 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
1478 |
+
for i, t in enumerate(timesteps):
|
1479 |
+
if self.interrupt:
|
1480 |
+
continue
|
1481 |
+
|
1482 |
+
latent_model_input = (
|
1483 |
+
torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
1484 |
+
)
|
1485 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
1486 |
+
|
1487 |
+
if i <= start_merge_step:
|
1488 |
+
current_prompt_embeds = torch.cat(
|
1489 |
+
[negative_prompt_embeds, prompt_embeds_text_only], dim=0
|
1490 |
+
)
|
1491 |
+
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds_text_only], dim=0)
|
1492 |
+
else:
|
1493 |
+
current_prompt_embeds = torch.cat(
|
1494 |
+
[negative_prompt_embeds, prompt_embeds], dim=0
|
1495 |
+
)
|
1496 |
+
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
|
1497 |
+
# predict the noise residual
|
1498 |
+
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
1499 |
+
# print(latent_model_input.shape)
|
1500 |
+
# print(t)
|
1501 |
+
# print(current_prompt_embeds.shape)
|
1502 |
+
# print(add_text_embeds.shape)
|
1503 |
+
# print(add_time_ids.shape)
|
1504 |
+
#zeros_matrix =
|
1505 |
+
#global_mask1024 = torch.cat([torch.randn(1, 1024, 1, 1, device=device) for random_number])
|
1506 |
+
#global_mask4096 =
|
1507 |
+
noise_pred = self.unet(
|
1508 |
+
latent_model_input,
|
1509 |
+
t,
|
1510 |
+
encoder_hidden_states=current_prompt_embeds,
|
1511 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
1512 |
+
added_cond_kwargs=added_cond_kwargs,
|
1513 |
+
return_dict=False,
|
1514 |
+
)[0]
|
1515 |
+
# print(noise_pred.shape)
|
1516 |
+
# perform guidance
|
1517 |
+
if do_classifier_free_guidance:
|
1518 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
1519 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
1520 |
+
|
1521 |
+
if do_classifier_free_guidance and guidance_rescale > 0.0:
|
1522 |
+
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
|
1523 |
+
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
|
1524 |
+
|
1525 |
+
# compute the previous noisy sample x_t -> x_t-1
|
1526 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
1527 |
+
|
1528 |
+
if callback_on_step_end is not None:
|
1529 |
+
callback_kwargs = {}
|
1530 |
+
for k in callback_on_step_end_tensor_inputs:
|
1531 |
+
callback_kwargs[k] = locals()[k]
|
1532 |
+
|
1533 |
+
ck_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
1534 |
+
|
1535 |
+
latents = callback_outputs.pop("latents", latents)
|
1536 |
+
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
1537 |
+
negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
|
1538 |
+
add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
|
1539 |
+
# negative_pooled_prompt_embeds = callback_outputs.pop(
|
1540 |
+
# "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
|
1541 |
+
# )
|
1542 |
+
# add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
|
1543 |
+
# negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
|
1544 |
+
|
1545 |
+
# call the callback, if provided
|
1546 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
1547 |
+
progress_bar.update()
|
1548 |
+
if callback is not None and i % callback_steps == 0:
|
1549 |
+
step_idx = i // getattr(self.scheduler, "order", 1)
|
1550 |
+
callback(step_idx, t, latents)
|
1551 |
+
|
1552 |
+
# make sure the VAE is in float32 mode, as it overflows in float16
|
1553 |
+
if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
|
1554 |
+
self.upcast_vae()
|
1555 |
+
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
|
1556 |
+
|
1557 |
+
if not output_type == "latent":
|
1558 |
+
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
1559 |
+
else:
|
1560 |
+
image = latents
|
1561 |
+
return StableDiffusionXLPipelineOutput(images=image)
|
1562 |
+
|
1563 |
+
# apply watermark if available
|
1564 |
+
# if self.watermark is not None:
|
1565 |
+
# image = self.watermark.apply_watermark(image)
|
1566 |
+
|
1567 |
+
image = self.image_processor.postprocess(image, output_type=output_type)
|
1568 |
+
|
1569 |
+
# Offload all models
|
1570 |
+
self.maybe_free_model_hooks()
|
1571 |
+
|
1572 |
+
if not return_dict:
|
1573 |
+
return (image,)
|
1574 |
+
|
1575 |
+
return StableDiffusionXLPipelineOutput(images=image)
|
1576 |
+
pipe2 = PhotoMakerStableDiffusionXLPipeline.from_pretrained(
|
1577 |
models_dict["Unstable"], torch_dtype=torch.float16, use_safetensors=use_safetensors)
|
1578 |
pipe2 = pipe2.to("cpu")
|
1579 |
pipe2.load_photomaker_adapter(
|