barreloflube
commited on
Commit
β’
5a59c13
1
Parent(s):
c7554bf
Refactor UI structure and import spaces module
Browse files- app2.py +721 -0
- src/tasks/images/init_sys.py +1 -1
app2.py
ADDED
@@ -0,0 +1,721 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Testing one file gradio app for zero gpu spaces not working as expected.
|
2 |
+
# Check here for the issue:
|
3 |
+
import gc
|
4 |
+
import json
|
5 |
+
import random
|
6 |
+
from typing import List, Optional
|
7 |
+
|
8 |
+
import spaces
|
9 |
+
import gradio as gr
|
10 |
+
from huggingface_hub import ModelCard
|
11 |
+
import torch
|
12 |
+
import numpy as np
|
13 |
+
from pydantic import BaseModel
|
14 |
+
from PIL import Image
|
15 |
+
from diffusers import (
|
16 |
+
FluxPipeline,
|
17 |
+
FluxImg2ImgPipeline,
|
18 |
+
FluxInpaintPipeline,
|
19 |
+
FluxControlNetPipeline,
|
20 |
+
StableDiffusionXLPipeline,
|
21 |
+
StableDiffusionXLImg2ImgPipeline,
|
22 |
+
StableDiffusionXLInpaintPipeline,
|
23 |
+
StableDiffusionXLControlNetPipeline,
|
24 |
+
StableDiffusionXLControlNetImg2ImgPipeline,
|
25 |
+
StableDiffusionXLControlNetInpaintPipeline,
|
26 |
+
AutoPipelineForText2Image,
|
27 |
+
AutoPipelineForImage2Image,
|
28 |
+
AutoPipelineForInpainting,
|
29 |
+
DiffusionPipeline,
|
30 |
+
AutoencoderKL,
|
31 |
+
FluxControlNetModel,
|
32 |
+
FluxMultiControlNetModel,
|
33 |
+
ControlNetModel,
|
34 |
+
)
|
35 |
+
from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
|
36 |
+
from huggingface_hub import hf_hub_download
|
37 |
+
from transformers import CLIPFeatureExtractor
|
38 |
+
from photomaker import FaceAnalysis2
|
39 |
+
from diffusers.schedulers import *
|
40 |
+
from huggingface_hub import hf_hub_download
|
41 |
+
from safetensors.torch import load_file
|
42 |
+
from controlnet_aux.processor import Processor
|
43 |
+
from photomaker import (
|
44 |
+
PhotoMakerStableDiffusionXLPipeline,
|
45 |
+
PhotoMakerStableDiffusionXLControlNetPipeline,
|
46 |
+
analyze_faces
|
47 |
+
)
|
48 |
+
from sd_embed.embedding_funcs import get_weighted_text_embeddings_sdxl, get_weighted_text_embeddings_flux1
|
49 |
+
|
50 |
+
|
51 |
+
# Initialize System
|
52 |
+
def load_sd():
|
53 |
+
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
54 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
55 |
+
|
56 |
+
# Models
|
57 |
+
models = [
|
58 |
+
{
|
59 |
+
"repo_id": "black-forest-labs/FLUX.1-dev",
|
60 |
+
"loader": "flux",
|
61 |
+
"compute_type": torch.bfloat16,
|
62 |
+
},
|
63 |
+
{
|
64 |
+
"repo_id": "SG161222/RealVisXL_V4.0",
|
65 |
+
"loader": "xl",
|
66 |
+
"compute_type": torch.float16,
|
67 |
+
}
|
68 |
+
]
|
69 |
+
|
70 |
+
for model in models:
|
71 |
+
try:
|
72 |
+
model["pipeline"] = AutoPipelineForText2Image.from_pretrained(
|
73 |
+
model['repo_id'],
|
74 |
+
torch_dtype = model['compute_type'],
|
75 |
+
safety_checker = None,
|
76 |
+
variant = "fp16"
|
77 |
+
).to(device)
|
78 |
+
model["pipeline"].enable_model_cpu_offload()
|
79 |
+
except:
|
80 |
+
model["pipeline"] = AutoPipelineForText2Image.from_pretrained(
|
81 |
+
model['repo_id'],
|
82 |
+
torch_dtype = model['compute_type'],
|
83 |
+
safety_checker = None
|
84 |
+
).to(device)
|
85 |
+
model["pipeline"].enable_model_cpu_offload()
|
86 |
+
|
87 |
+
|
88 |
+
# VAE n Refiner
|
89 |
+
sdxl_vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to(device)
|
90 |
+
refiner = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0", vae=sdxl_vae, torch_dtype=torch.float16, use_safetensors=True, variant="fp16").to(device)
|
91 |
+
refiner.enable_model_cpu_offload()
|
92 |
+
|
93 |
+
|
94 |
+
# Safety Checker
|
95 |
+
safety_checker = StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-safety-checker").to(device)
|
96 |
+
feature_extractor = CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32", from_pt=True)
|
97 |
+
|
98 |
+
|
99 |
+
# Controlnets
|
100 |
+
controlnet_models = [
|
101 |
+
{
|
102 |
+
"repo_id": "xinsir/controlnet-depth-sdxl-1.0",
|
103 |
+
"name": "depth_xl",
|
104 |
+
"layers": ["depth"],
|
105 |
+
"loader": "xl",
|
106 |
+
"compute_type": torch.float16,
|
107 |
+
},
|
108 |
+
{
|
109 |
+
"repo_id": "xinsir/controlnet-canny-sdxl-1.0",
|
110 |
+
"name": "canny_xl",
|
111 |
+
"layers": ["canny"],
|
112 |
+
"loader": "xl",
|
113 |
+
"compute_type": torch.float16,
|
114 |
+
},
|
115 |
+
{
|
116 |
+
"repo_id": "xinsir/controlnet-openpose-sdxl-1.0",
|
117 |
+
"name": "openpose_xl",
|
118 |
+
"layers": ["pose"],
|
119 |
+
"loader": "xl",
|
120 |
+
"compute_type": torch.float16,
|
121 |
+
},
|
122 |
+
{
|
123 |
+
"repo_id": "xinsir/controlnet-scribble-sdxl-1.0",
|
124 |
+
"name": "scribble_xl",
|
125 |
+
"layers": ["scribble"],
|
126 |
+
"loader": "xl",
|
127 |
+
"compute_type": torch.float16,
|
128 |
+
},
|
129 |
+
{
|
130 |
+
"repo_id": "Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro",
|
131 |
+
"name": "flux1_union_pro",
|
132 |
+
"layers": ["canny_fl", "tile_fl", "depth_fl", "blur_fl", "pose_fl", "gray_fl", "low_quality_fl"],
|
133 |
+
"loader": "flux-multi",
|
134 |
+
"compute_type": torch.bfloat16,
|
135 |
+
}
|
136 |
+
]
|
137 |
+
|
138 |
+
for controlnet in controlnet_models:
|
139 |
+
if controlnet["loader"] == "xl":
|
140 |
+
controlnet["controlnet"] = ControlNetModel.from_pretrained(
|
141 |
+
controlnet["repo_id"],
|
142 |
+
torch_dtype = controlnet['compute_type']
|
143 |
+
).to(device)
|
144 |
+
elif controlnet["loader"] == "flux-multi":
|
145 |
+
controlnet["controlnet"] = FluxMultiControlNetModel([FluxControlNetModel.from_pretrained(
|
146 |
+
controlnet["repo_id"],
|
147 |
+
torch_dtype = controlnet['compute_type']
|
148 |
+
).to(device)])
|
149 |
+
#TODO: Add support for flux only controlnet
|
150 |
+
|
151 |
+
|
152 |
+
# Face Detection (for PhotoMaker)
|
153 |
+
face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
|
154 |
+
face_detector.prepare(ctx_id=0, det_size=(640, 640))
|
155 |
+
|
156 |
+
|
157 |
+
# PhotoMaker V2 (for SDXL only)
|
158 |
+
photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker-V2", filename="photomaker-v2.bin", repo_type="model")
|
159 |
+
|
160 |
+
return device, models, sdxl_vae, refiner, safety_checker, feature_extractor, controlnet_models, face_detector, photomaker_ckpt
|
161 |
+
|
162 |
+
|
163 |
+
device, models, sdxl_vae, refiner, safety_checker, feature_extractor, controlnet_models, face_detector, photomaker_ckpt = load_sd()
|
164 |
+
|
165 |
+
|
166 |
+
# Models
|
167 |
+
class ControlNetReq(BaseModel):
|
168 |
+
controlnets: List[str] # ["canny", "tile", "depth"]
|
169 |
+
control_images: List[Image.Image]
|
170 |
+
controlnet_conditioning_scale: List[float]
|
171 |
+
|
172 |
+
class Config:
|
173 |
+
arbitrary_types_allowed=True
|
174 |
+
|
175 |
+
|
176 |
+
class SDReq(BaseModel):
|
177 |
+
model: str = ""
|
178 |
+
prompt: str = ""
|
179 |
+
negative_prompt: Optional[str] = "black-forest-labs/FLUX.1-dev"
|
180 |
+
fast_generation: Optional[bool] = True
|
181 |
+
loras: Optional[list] = []
|
182 |
+
embeddings: Optional[list] = []
|
183 |
+
resize_mode: Optional[str] = "resize_and_fill" # resize_only, crop_and_resize, resize_and_fill
|
184 |
+
scheduler: Optional[str] = "euler_fl"
|
185 |
+
height: int = 1024
|
186 |
+
width: int = 1024
|
187 |
+
num_images_per_prompt: int = 1
|
188 |
+
num_inference_steps: int = 8
|
189 |
+
guidance_scale: float = 3.5
|
190 |
+
seed: Optional[int] = 0
|
191 |
+
refiner: bool = False
|
192 |
+
vae: bool = True
|
193 |
+
controlnet_config: Optional[ControlNetReq] = None
|
194 |
+
photomaker_images: Optional[List[Image.Image]] = None
|
195 |
+
|
196 |
+
class Config:
|
197 |
+
arbitrary_types_allowed=True
|
198 |
+
|
199 |
+
|
200 |
+
class SDImg2ImgReq(SDReq):
|
201 |
+
image: Image.Image
|
202 |
+
strength: float = 1.0
|
203 |
+
|
204 |
+
class Config:
|
205 |
+
arbitrary_types_allowed=True
|
206 |
+
|
207 |
+
|
208 |
+
class SDInpaintReq(SDImg2ImgReq):
|
209 |
+
mask_image: Image.Image
|
210 |
+
|
211 |
+
class Config:
|
212 |
+
arbitrary_types_allowed=True
|
213 |
+
|
214 |
+
|
215 |
+
# Helper functions
|
216 |
+
def get_controlnet(controlnet_config: ControlNetReq):
|
217 |
+
control_mode = []
|
218 |
+
controlnet = []
|
219 |
+
|
220 |
+
for m in controlnet_models:
|
221 |
+
for c in controlnet_config.controlnets:
|
222 |
+
if c in m["layers"]:
|
223 |
+
control_mode.append(m["layers"].index(c))
|
224 |
+
controlnet.append(m["controlnet"])
|
225 |
+
|
226 |
+
return controlnet, control_mode
|
227 |
+
|
228 |
+
|
229 |
+
def get_pipe(request: SDReq | SDImg2ImgReq | SDInpaintReq):
|
230 |
+
for m in models:
|
231 |
+
if m["repo_id"] == request.model:
|
232 |
+
pipeline = m['pipeline']
|
233 |
+
controlnet, control_mode = get_controlnet(request.controlnet_config) if request.controlnet_config else (None, None)
|
234 |
+
|
235 |
+
pipe_args = {
|
236 |
+
"pipeline": pipeline,
|
237 |
+
"control_mode": control_mode,
|
238 |
+
}
|
239 |
+
if request.controlnet_config:
|
240 |
+
pipe_args["controlnet"] = controlnet
|
241 |
+
|
242 |
+
if not request.photomaker_images:
|
243 |
+
if isinstance(request, SDReq):
|
244 |
+
pipe_args['pipeline'] = AutoPipelineForText2Image.from_pipe(**pipe_args)
|
245 |
+
elif isinstance(request, SDImg2ImgReq):
|
246 |
+
pipe_args['pipeline'] = AutoPipelineForImage2Image.from_pipe(**pipe_args)
|
247 |
+
elif isinstance(request, SDInpaintReq):
|
248 |
+
pipe_args['pipeline'] = AutoPipelineForInpainting.from_pipe(**pipe_args)
|
249 |
+
else:
|
250 |
+
raise ValueError(f"Unknown request type: {type(request)}")
|
251 |
+
elif isinstance(request, any([PhotoMakerStableDiffusionXLPipeline, PhotoMakerStableDiffusionXLControlNetPipeline])):
|
252 |
+
if request.controlnet_config:
|
253 |
+
pipe_args['pipeline'] = PhotoMakerStableDiffusionXLControlNetPipeline.from_pipe(**pipe_args)
|
254 |
+
else:
|
255 |
+
pipe_args['pipeline'] = PhotoMakerStableDiffusionXLPipeline.from_pipe(**pipe_args)
|
256 |
+
else:
|
257 |
+
raise ValueError(f"Invalid request type: {type(request)}")
|
258 |
+
|
259 |
+
return pipe_args
|
260 |
+
|
261 |
+
|
262 |
+
def load_scheduler(pipeline, scheduler):
|
263 |
+
schedulers = {
|
264 |
+
"dpmpp_2m": (DPMSolverMultistepScheduler, {}),
|
265 |
+
"dpmpp_2m_k": (DPMSolverMultistepScheduler, {"use_karras_sigmas": True}),
|
266 |
+
"dpmpp_2m_sde": (DPMSolverMultistepScheduler, {"algorithm_type": "sde-dpmsolver++"}),
|
267 |
+
"dpmpp_2m_sde_k": (DPMSolverMultistepScheduler, {"algorithm_type": "sde-dpmsolver++", "use_karras_sigmas": True}),
|
268 |
+
"dpmpp_sde": (DPMSolverSinglestepScheduler, {}),
|
269 |
+
"dpmpp_sde_k": (DPMSolverSinglestepScheduler, {"use_karras_sigmas": True}),
|
270 |
+
"dpm2": (KDPM2DiscreteScheduler, {}),
|
271 |
+
"dpm2_k": (KDPM2DiscreteScheduler, {"use_karras_sigmas": True}),
|
272 |
+
"dpm2_a": (KDPM2AncestralDiscreteScheduler, {}),
|
273 |
+
"dpm2_a_k": (KDPM2AncestralDiscreteScheduler, {"use_karras_sigmas": True}),
|
274 |
+
"euler": (EulerDiscreteScheduler, {}),
|
275 |
+
"euler_a": (EulerAncestralDiscreteScheduler, {}),
|
276 |
+
"heun": (HeunDiscreteScheduler, {}),
|
277 |
+
"lms": (LMSDiscreteScheduler, {}),
|
278 |
+
"lms_k": (LMSDiscreteScheduler, {"use_karras_sigmas": True}),
|
279 |
+
"deis": (DEISMultistepScheduler, {}),
|
280 |
+
"unipc": (UniPCMultistepScheduler, {}),
|
281 |
+
"fm_euler": (FlowMatchEulerDiscreteScheduler, {}),
|
282 |
+
}
|
283 |
+
scheduler_class, kwargs = schedulers.get(scheduler, (None, {}))
|
284 |
+
|
285 |
+
if scheduler_class is not None:
|
286 |
+
scheduler = scheduler_class.from_config(pipeline.scheduler.config, **kwargs)
|
287 |
+
else:
|
288 |
+
raise ValueError(f"Unknown scheduler: {scheduler}")
|
289 |
+
|
290 |
+
return scheduler
|
291 |
+
|
292 |
+
|
293 |
+
def load_loras(pipeline, loras, fast_generation):
|
294 |
+
for i, lora in enumerate(loras):
|
295 |
+
pipeline.load_lora_weights(lora['repo_id'], adapter_name=f"lora_{i}")
|
296 |
+
adapter_names = [f"lora_{i}" for i in range(len(loras))]
|
297 |
+
adapter_weights = [lora['weight'] for lora in loras]
|
298 |
+
|
299 |
+
if fast_generation:
|
300 |
+
hyper_lora = hf_hub_download(
|
301 |
+
"ByteDance/Hyper-SD",
|
302 |
+
"Hyper-FLUX.1-dev-8steps-lora.safetensors" if isinstance(pipeline, FluxPipeline) else "Hyper-SDXL-2steps-lora.safetensors"
|
303 |
+
)
|
304 |
+
hyper_weight = 0.125 if isinstance(pipeline, FluxPipeline) else 1.0
|
305 |
+
pipeline.load_lora_weights(hyper_lora, adapter_name="hyper_lora")
|
306 |
+
adapter_names.append("hyper_lora")
|
307 |
+
adapter_weights.append(hyper_weight)
|
308 |
+
|
309 |
+
pipeline.set_adapters(adapter_names, adapter_weights)
|
310 |
+
|
311 |
+
|
312 |
+
def load_xl_embeddings(pipeline, embeddings):
|
313 |
+
for embedding in embeddings:
|
314 |
+
state_dict = load_file(hf_hub_download(embedding['repo_id']))
|
315 |
+
pipeline.load_textual_inversion(state_dict['clip_g'], token=embedding['token'], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2)
|
316 |
+
pipeline.load_textual_inversion(state_dict["clip_l"], token=embedding['token'], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer)
|
317 |
+
|
318 |
+
|
319 |
+
def resize_images(images: List[Image.Image], height: int, width: int, resize_mode: str):
|
320 |
+
for image in images:
|
321 |
+
if resize_mode == "resize_only":
|
322 |
+
image = image.resize((width, height))
|
323 |
+
elif resize_mode == "crop_and_resize":
|
324 |
+
image = image.crop((0, 0, width, height))
|
325 |
+
elif resize_mode == "resize_and_fill":
|
326 |
+
image = image.resize((width, height), Image.Resampling.LANCZOS)
|
327 |
+
|
328 |
+
return images
|
329 |
+
|
330 |
+
|
331 |
+
def get_controlnet_images(controlnets: List[str], control_images: List[Image.Image], height: int, width: int, resize_mode: str):
|
332 |
+
response_images = []
|
333 |
+
control_images = resize_images(control_images, height, width, resize_mode)
|
334 |
+
for controlnet, image in zip(controlnets, control_images):
|
335 |
+
if controlnet == "canny" or controlnet == "canny_xs" or controlnet == "canny_fl":
|
336 |
+
processor = Processor('canny')
|
337 |
+
elif controlnet == "depth" or controlnet == "depth_xs" or controlnet == "depth_fl":
|
338 |
+
processor = Processor('depth_midas')
|
339 |
+
elif controlnet == "pose" or controlnet == "pose_fl":
|
340 |
+
processor = Processor('openpose_full')
|
341 |
+
elif controlnet == "scribble":
|
342 |
+
processor = Processor('scribble')
|
343 |
+
else:
|
344 |
+
raise ValueError(f"Invalid Controlnet: {controlnet}")
|
345 |
+
|
346 |
+
response_images.append(processor(image, to_pil=True))
|
347 |
+
|
348 |
+
return response_images
|
349 |
+
|
350 |
+
|
351 |
+
def check_image_safety(images: List[Image.Image]):
|
352 |
+
safety_checker_input = feature_extractor(images, return_tensors="pt").to("cuda")
|
353 |
+
has_nsfw_concepts = safety_checker(
|
354 |
+
images=[images],
|
355 |
+
clip_input=safety_checker_input.pixel_values.to("cuda"),
|
356 |
+
)
|
357 |
+
|
358 |
+
return has_nsfw_concepts[1]
|
359 |
+
|
360 |
+
|
361 |
+
def get_prompt_attention(pipeline, prompt, negative_prompt):
|
362 |
+
if isinstance(pipeline, (FluxPipeline, FluxImg2ImgPipeline, FluxInpaintPipeline, FluxControlNetPipeline)):
|
363 |
+
prompt_embeds, pooled_prompt_embeds = get_weighted_text_embeddings_flux1(pipeline, prompt)
|
364 |
+
return prompt_embeds, None, pooled_prompt_embeds, None
|
365 |
+
elif isinstance(pipeline, StableDiffusionXLPipeline):
|
366 |
+
prompt_embeds, prompt_neg_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds = get_weighted_text_embeddings_sdxl(pipeline, prompt, negative_prompt)
|
367 |
+
return prompt_embeds, prompt_neg_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
|
368 |
+
else:
|
369 |
+
raise ValueError(f"Invalid pipeline type: {type(pipeline)}")
|
370 |
+
|
371 |
+
|
372 |
+
def get_photomaker_images(photomaker_images: List[Image.Image], height: int, width: int, resize_mode: str):
|
373 |
+
image_input_ids = []
|
374 |
+
image_id_embeds = []
|
375 |
+
photomaker_images = resize_images(photomaker_images, height, width, resize_mode)
|
376 |
+
|
377 |
+
for image in photomaker_images:
|
378 |
+
image_input_ids.append(img)
|
379 |
+
img = np.array(image)[:, :, ::-1]
|
380 |
+
faces = analyze_faces(face_detector, image)
|
381 |
+
if len(faces) > 0:
|
382 |
+
image_id_embeds.append(torch.from_numpy(faces[0]['embeddings']))
|
383 |
+
else:
|
384 |
+
raise ValueError("No face detected in the image")
|
385 |
+
|
386 |
+
return image_input_ids, image_id_embeds
|
387 |
+
|
388 |
+
|
389 |
+
def cleanup(pipeline, loras = None, embeddings = None):
|
390 |
+
if loras:
|
391 |
+
pipeline.disable_lora()
|
392 |
+
pipeline.unload_lora_weights()
|
393 |
+
if embeddings:
|
394 |
+
pipeline.unload_textual_inversion()
|
395 |
+
gc.collect()
|
396 |
+
torch.cuda.empty_cache()
|
397 |
+
|
398 |
+
|
399 |
+
# Gen function
|
400 |
+
def gen_img(
|
401 |
+
request: SDReq | SDImg2ImgReq | SDInpaintReq
|
402 |
+
):
|
403 |
+
pipeline_args = get_pipe(request)
|
404 |
+
pipeline = pipeline_args['pipeline']
|
405 |
+
try:
|
406 |
+
pipeline.scheduler = load_scheduler(pipeline, request.scheduler)
|
407 |
+
|
408 |
+
load_loras(pipeline, request.loras, request.fast_generation)
|
409 |
+
load_xl_embeddings(pipeline, request.embeddings)
|
410 |
+
|
411 |
+
control_images = get_controlnet_images(request.controlnet_config.controlnets, request.controlnet_config.control_images, request.height, request.width, request.resize_mode) if request.controlnet_config else None
|
412 |
+
photomaker_images, photomaker_id_embeds = get_photomaker_images(request.photomaker_images, request.height, request.width) if request.photomaker_images else (None, None)
|
413 |
+
|
414 |
+
positive_prompt_embeds, negative_prompt_embeds, positive_prompt_pooled, negative_prompt_pooled = get_prompt_attention(pipeline, request.prompt, request.negative_prompt)
|
415 |
+
|
416 |
+
# Common args
|
417 |
+
args = {
|
418 |
+
'prompt_embeds': positive_prompt_embeds,
|
419 |
+
'pooled_prompt_embeds': positive_prompt_pooled,
|
420 |
+
'height': request.height,
|
421 |
+
'width': request.width,
|
422 |
+
'num_images_per_prompt': request.num_images_per_prompt,
|
423 |
+
'num_inference_steps': request.num_inference_steps,
|
424 |
+
'guidance_scale': request.guidance_scale,
|
425 |
+
'generator': [torch.Generator(device=device).manual_seed(request.seed + i) if not request.seed is any([None, 0, -1]) else torch.Generator(device=device).manual_seed(random.randint(0, 2**32 - 1)) for i in range(request.num_images_per_prompt)],
|
426 |
+
}
|
427 |
+
|
428 |
+
if isinstance(pipeline, any([StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline, StableDiffusionXLInpaintPipeline,
|
429 |
+
StableDiffusionXLControlNetPipeline, StableDiffusionXLControlNetImg2ImgPipeline, StableDiffusionXLControlNetInpaintPipeline])):
|
430 |
+
args['clip_skip'] = request.clip_skip
|
431 |
+
args['negative_prompt_embeds'] = negative_prompt_embeds
|
432 |
+
args['negative_pooled_prompt_embeds'] = negative_prompt_pooled
|
433 |
+
|
434 |
+
if isinstance(pipeline, FluxControlNetPipeline) and request.controlnet_config:
|
435 |
+
args['control_mode'] = pipeline_args['control_mode']
|
436 |
+
args['control_image'] = control_images
|
437 |
+
args['controlnet_conditioning_scale'] = request.controlnet_conditioning_scale
|
438 |
+
|
439 |
+
if not isinstance(pipeline, FluxControlNetPipeline) and request.controlnet_config:
|
440 |
+
args['controlnet_conditioning_scale'] = request.controlnet_conditioning_scale
|
441 |
+
|
442 |
+
if isinstance(request, SDReq):
|
443 |
+
args['image'] = control_images
|
444 |
+
elif isinstance(request, (SDImg2ImgReq, SDInpaintReq)):
|
445 |
+
args['control_image'] = control_images
|
446 |
+
|
447 |
+
if request.photomaker_images and isinstance(pipeline, any([PhotoMakerStableDiffusionXLPipeline, PhotoMakerStableDiffusionXLControlNetPipeline])):
|
448 |
+
args['input_id_images'] = photomaker_images
|
449 |
+
args['input_id_embeds'] = photomaker_id_embeds
|
450 |
+
args['start_merge_step'] = 10
|
451 |
+
|
452 |
+
if isinstance(request, SDImg2ImgReq):
|
453 |
+
args['image'] = resize_images([request.image], request.height, request.width, request.resize_mode)
|
454 |
+
args['strength'] = request.strength
|
455 |
+
elif isinstance(request, SDInpaintReq):
|
456 |
+
args['image'] = resize_images([request.image], request.height, request.width, request.resize_mode)
|
457 |
+
args['mask_image'] = resize_images([request.mask_image], request.height, request.width, request.resize_mode)
|
458 |
+
args['strength'] = request.strength
|
459 |
+
|
460 |
+
images = pipeline(**args).images
|
461 |
+
|
462 |
+
if request.refiner:
|
463 |
+
images = refiner(
|
464 |
+
prompt=request.prompt,
|
465 |
+
num_inference_steps=40,
|
466 |
+
denoising_start=0.7,
|
467 |
+
image=images.images
|
468 |
+
).images
|
469 |
+
|
470 |
+
cleanup(pipeline, request.loras, request.embeddings)
|
471 |
+
|
472 |
+
return images
|
473 |
+
except Exception as e:
|
474 |
+
cleanup(pipeline, request.loras, request.embeddings)
|
475 |
+
raise ValueError(f"Error generating image: {e}") from e
|
476 |
+
|
477 |
+
|
478 |
+
# CSS
|
479 |
+
css = """
|
480 |
+
@import url('https://fonts.googleapis.com/css2?family=Poppins:wght@300;400;600&display=swap');
|
481 |
+
body {
|
482 |
+
font-family: 'Poppins', sans-serif !important;
|
483 |
+
}
|
484 |
+
.center-content {
|
485 |
+
text-align: center;
|
486 |
+
max-width: 600px;
|
487 |
+
margin: 0 auto;
|
488 |
+
padding: 20px;
|
489 |
+
}
|
490 |
+
.center-content h1 {
|
491 |
+
font-weight: 600;
|
492 |
+
margin-bottom: 1rem;
|
493 |
+
}
|
494 |
+
.center-content p {
|
495 |
+
margin-bottom: 1.5rem;
|
496 |
+
}
|
497 |
+
"""
|
498 |
+
|
499 |
+
|
500 |
+
flux_models = ["black-forest-labs/FLUX.1-dev"]
|
501 |
+
with open("data/images/loras/flux.json", "r") as f:
|
502 |
+
loras = json.load(f)
|
503 |
+
|
504 |
+
|
505 |
+
# Main Gradio app
|
506 |
+
with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
507 |
+
# Header
|
508 |
+
with gr.Column(elem_classes="center-content"):
|
509 |
+
gr.Markdown("""
|
510 |
+
# π AAI: All AI
|
511 |
+
Unleash your creativity with our multi-modal AI platform.
|
512 |
+
[![Sync code to HF Space](https://github.com/mantrakp04/aai/actions/workflows/hf-space.yml/badge.svg)](https://github.com/mantrakp04/aai/actions/workflows/hf-space.yml)
|
513 |
+
""")
|
514 |
+
|
515 |
+
# Tabs
|
516 |
+
with gr.Tabs():
|
517 |
+
with gr.Tab(label="πΌοΈ Image"):
|
518 |
+
with gr.Tabs():
|
519 |
+
with gr.Tab("Flux"):
|
520 |
+
"""
|
521 |
+
Create the image tab for Generative Image Generation Models
|
522 |
+
|
523 |
+
Args:
|
524 |
+
models: list
|
525 |
+
A list containing the models repository paths
|
526 |
+
gap_iol, gap_la, gap_le, gap_eio, gap_io: Optional[List[dict]]
|
527 |
+
A list of dictionaries containing the title and component for the custom gradio component
|
528 |
+
Example:
|
529 |
+
def gr_comp():
|
530 |
+
gr.Label("Hello World")
|
531 |
+
|
532 |
+
[
|
533 |
+
{
|
534 |
+
'title': "Title",
|
535 |
+
'component': gr_comp()
|
536 |
+
}
|
537 |
+
]
|
538 |
+
loras: list
|
539 |
+
A list of dictionaries containing the image and title for the Loras Gallery
|
540 |
+
Generally a loaded json file from the data folder
|
541 |
+
|
542 |
+
"""
|
543 |
+
def process_gaps(gaps: List[dict]):
|
544 |
+
for gap in gaps:
|
545 |
+
with gr.Accordion(gap['title']):
|
546 |
+
gap['component']
|
547 |
+
|
548 |
+
|
549 |
+
with gr.Row():
|
550 |
+
with gr.Column():
|
551 |
+
with gr.Group() as image_options:
|
552 |
+
model = gr.Dropdown(label="Models", choices=flux_models, value=flux_models[0], interactive=True)
|
553 |
+
prompt = gr.Textbox(lines=5, label="Prompt")
|
554 |
+
negative_prompt = gr.Textbox(label="Negative Prompt")
|
555 |
+
fast_generation = gr.Checkbox(label="Fast Generation (Hyper-SD) π§ͺ")
|
556 |
+
|
557 |
+
|
558 |
+
with gr.Accordion("Loras", open=True): # Lora Gallery
|
559 |
+
lora_gallery = gr.Gallery(
|
560 |
+
label="Gallery",
|
561 |
+
value=[(lora['image'], lora['title']) for lora in loras],
|
562 |
+
allow_preview=False,
|
563 |
+
columns=[3],
|
564 |
+
type="pil"
|
565 |
+
)
|
566 |
+
|
567 |
+
with gr.Group():
|
568 |
+
with gr.Column():
|
569 |
+
with gr.Row():
|
570 |
+
custom_lora = gr.Textbox(label="Custom Lora", info="Enter a Huggingface repo path")
|
571 |
+
selected_lora = gr.Textbox(label="Selected Lora", info="Choose from the gallery or enter a custom LoRA")
|
572 |
+
|
573 |
+
custom_lora_info = gr.HTML(visible=False)
|
574 |
+
add_lora = gr.Button(value="Add LoRA")
|
575 |
+
|
576 |
+
enabled_loras = gr.State(value=[])
|
577 |
+
with gr.Group():
|
578 |
+
with gr.Row():
|
579 |
+
for i in range(6): # only support max 6 loras due to inference time
|
580 |
+
with gr.Column():
|
581 |
+
with gr.Column(scale=2):
|
582 |
+
globals()[f"lora_slider_{i}"] = gr.Slider(label=f"LoRA {i+1}", minimum=0, maximum=1, step=0.01, value=0.8, visible=False, interactive=True)
|
583 |
+
with gr.Column():
|
584 |
+
globals()[f"lora_remove_{i}"] = gr.Button(value="Remove LoRA", visible=False)
|
585 |
+
|
586 |
+
|
587 |
+
with gr.Accordion("Embeddings", open=False): # Embeddings
|
588 |
+
gr.Label("To be implemented")
|
589 |
+
|
590 |
+
|
591 |
+
with gr.Accordion("Image Options"): # Image Options
|
592 |
+
with gr.Tabs():
|
593 |
+
image_options = {
|
594 |
+
"img2img": "Upload Image",
|
595 |
+
"inpaint": "Upload Image",
|
596 |
+
"canny": "Upload Image",
|
597 |
+
"pose": "Upload Image",
|
598 |
+
"depth": "Upload Image",
|
599 |
+
}
|
600 |
+
|
601 |
+
for image_option, label in image_options.items():
|
602 |
+
with gr.Tab(image_option):
|
603 |
+
if not image_option in ['inpaint', 'scribble']:
|
604 |
+
globals()[f"{image_option}_image"] = gr.Image(label=label, type="pil")
|
605 |
+
elif image_option in ['inpaint', 'scribble']:
|
606 |
+
globals()[f"{image_option}_image"] = gr.ImageEditor(
|
607 |
+
label=label,
|
608 |
+
image_mode='RGB',
|
609 |
+
layers=False,
|
610 |
+
brush=gr.Brush(colors=["#FFFFFF"], color_mode="fixed") if image_option == 'inpaint' else gr.Brush(),
|
611 |
+
interactive=True,
|
612 |
+
type="pil",
|
613 |
+
)
|
614 |
+
|
615 |
+
# Image Strength (Co-relates to controlnet strength, strength for img2img n inpaint)
|
616 |
+
globals()[f"{image_option}_strength"] = gr.Slider(label="Strength", minimum=0, maximum=1, step=0.01, value=1.0, interactive=True)
|
617 |
+
|
618 |
+
resize_mode = gr.Radio(
|
619 |
+
label="Resize Mode",
|
620 |
+
choices=["crop and resize", "resize only", "resize and fill"],
|
621 |
+
value="resize and fill",
|
622 |
+
interactive=True
|
623 |
+
)
|
624 |
+
|
625 |
+
|
626 |
+
with gr.Column():
|
627 |
+
with gr.Group():
|
628 |
+
output_images = gr.Gallery(
|
629 |
+
label="Output Images",
|
630 |
+
value=[],
|
631 |
+
allow_preview=True,
|
632 |
+
type="pil",
|
633 |
+
interactive=False,
|
634 |
+
)
|
635 |
+
generate_images = gr.Button(value="Generate Images", variant="primary")
|
636 |
+
|
637 |
+
with gr.Accordion("Advance Settings", open=True):
|
638 |
+
with gr.Row():
|
639 |
+
scheduler = gr.Dropdown(
|
640 |
+
label="Scheduler",
|
641 |
+
choices = [
|
642 |
+
"fm_euler"
|
643 |
+
],
|
644 |
+
value="fm_euler",
|
645 |
+
interactive=True
|
646 |
+
)
|
647 |
+
|
648 |
+
with gr.Row():
|
649 |
+
for column in range(2):
|
650 |
+
with gr.Column():
|
651 |
+
options = [
|
652 |
+
("Height", "image_height", 64, 1024, 64, 1024, True),
|
653 |
+
("Width", "image_width", 64, 1024, 64, 1024, True),
|
654 |
+
("Num Images Per Prompt", "image_num_images_per_prompt", 1, 4, 1, 1, True),
|
655 |
+
("Num Inference Steps", "image_num_inference_steps", 1, 100, 1, 20, True),
|
656 |
+
("Clip Skip", "image_clip_skip", 0, 2, 1, 2, False),
|
657 |
+
("Guidance Scale", "image_guidance_scale", 0, 20, 0.5, 3.5, True),
|
658 |
+
("Seed", "image_seed", 0, 100000, 1, random.randint(0, 100000), True),
|
659 |
+
]
|
660 |
+
for label, var_name, min_val, max_val, step, value, visible in options[column::2]:
|
661 |
+
globals()[var_name] = gr.Slider(label=label, minimum=min_val, maximum=max_val, step=step, value=value, visible=visible, interactive=True)
|
662 |
+
|
663 |
+
with gr.Row():
|
664 |
+
refiner = gr.Checkbox(
|
665 |
+
label="Refiner π§ͺ",
|
666 |
+
value=False,
|
667 |
+
)
|
668 |
+
vae = gr.Checkbox(
|
669 |
+
label="VAE",
|
670 |
+
value=True,
|
671 |
+
)
|
672 |
+
|
673 |
+
|
674 |
+
# Events
|
675 |
+
# Base Options
|
676 |
+
fast_generation.change(update_fast_generation, [model, fast_generation], [image_guidance_scale, image_num_inference_steps]) # Fast Generation # type: ignore
|
677 |
+
|
678 |
+
|
679 |
+
# Lora Gallery
|
680 |
+
lora_gallery.select(selected_lora_from_gallery, None, selected_lora)
|
681 |
+
custom_lora.change(update_selected_lora, custom_lora, [custom_lora, selected_lora])
|
682 |
+
add_lora.click(add_to_enabled_loras, [model, selected_lora, enabled_loras], [selected_lora, custom_lora_info, enabled_loras])
|
683 |
+
enabled_loras.change(update_lora_sliders, enabled_loras, [lora_slider_0, lora_slider_1, lora_slider_2, lora_slider_3, lora_slider_4, lora_slider_5, lora_remove_0, lora_remove_1, lora_remove_2, lora_remove_3, lora_remove_4, lora_remove_5]) # type: ignore
|
684 |
+
|
685 |
+
for i in range(6):
|
686 |
+
globals()[f"lora_remove_{i}"].click(
|
687 |
+
lambda enabled_loras, index=i: remove_from_enabled_loras(enabled_loras, index),
|
688 |
+
[enabled_loras],
|
689 |
+
[enabled_loras]
|
690 |
+
)
|
691 |
+
|
692 |
+
|
693 |
+
# Generate Image
|
694 |
+
generate_images.click(
|
695 |
+
generate_image, # type: ignore
|
696 |
+
[
|
697 |
+
model, prompt, negative_prompt, fast_generation, enabled_loras,
|
698 |
+
lora_slider_0, lora_slider_1, lora_slider_2, lora_slider_3, lora_slider_4, lora_slider_5, # type: ignore
|
699 |
+
img2img_image, inpaint_image, canny_image, pose_image, depth_image, # type: ignore
|
700 |
+
img2img_strength, inpaint_strength, canny_strength, pose_strength, depth_strength, # type: ignore
|
701 |
+
resize_mode,
|
702 |
+
scheduler, image_height, image_width, image_num_images_per_prompt, # type: ignore
|
703 |
+
image_num_inference_steps, image_guidance_scale, image_seed, # type: ignore
|
704 |
+
refiner, vae
|
705 |
+
],
|
706 |
+
[output_images]
|
707 |
+
)
|
708 |
+
with gr.Tab("SDXL"):
|
709 |
+
gr.Label("To be implemented")
|
710 |
+
with gr.Tab(label="π΅ Audio"):
|
711 |
+
gr.Label("Coming soon!")
|
712 |
+
with gr.Tab(label="π¬ Video"):
|
713 |
+
gr.Label("Coming soon!")
|
714 |
+
with gr.Tab(label="π Text"):
|
715 |
+
gr.Label("Coming soon!")
|
716 |
+
|
717 |
+
|
718 |
+
demo.launch(
|
719 |
+
share=False,
|
720 |
+
debug=True,
|
721 |
+
)
|
src/tasks/images/init_sys.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
|
2 |
import torch
|
3 |
from diffusers import (
|
4 |
DiffusionPipeline,
|
|
|
1 |
+
import spaces
|
2 |
import torch
|
3 |
from diffusers import (
|
4 |
DiffusionPipeline,
|