Spaces:
Running
Running
File size: 24,591 Bytes
78506fe c53d44b 95bdd33 0b30521 78506fe 0b30521 0b744ad 0b30521 78506fe 2dd5330 0b30521 2dd5330 78506fe e33e382 c2401c0 e33e382 78506fe 95bdd33 78506fe 95bdd33 08456a1 95bdd33 78506fe 95bdd33 78506fe 95bdd33 78506fe 95bdd33 c53d44b 78506fe 95bdd33 78506fe d36da76 95bdd33 78506fe 95bdd33 78506fe 95bdd33 78506fe c2401c0 78506fe 5edca44 c2401c0 77987e2 c2401c0 77987e2 c2401c0 e2eb8eb e33e382 e2eb8eb e33e382 e2eb8eb 78506fe e33e382 e2eb8eb 45523be e2eb8eb 78506fe |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 |
import gradio as gr
import random
import json
import os
import re
from datetime import datetime
from huggingface_hub import InferenceClient
import subprocess
import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
# Initialize Florence model
device = "cuda" if torch.cuda.is_available() else "cpu"
florence_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval()
florence_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True)
# Florence caption function
def florence_caption(image):
if not isinstance(image, Image.Image):
image = Image.fromarray(image)
inputs = florence_processor(text="<MORE_DETAILED_CAPTION>", images=image, return_tensors="pt").to(device)
generated_ids = florence_model.generate(
input_ids=inputs["input_ids"],
pixel_values=inputs["pixel_values"],
max_new_tokens=1024,
early_stopping=False,
do_sample=False,
num_beams=3,
)
generated_text = florence_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
parsed_answer = florence_processor.post_process_generation(
generated_text,
task="<MORE_DETAILED_CAPTION>",
image_size=(image.width, image.height)
)
return parsed_answer["<MORE_DETAILED_CAPTION>"]
# Load JSON files
def load_json_file(file_name):
file_path = os.path.join("data", file_name)
with open(file_path, "r") as file:
return json.load(file)
ARTFORM = load_json_file("artform.json")
PHOTO_TYPE = load_json_file("photo_type.json")
BODY_TYPES = load_json_file("body_types.json")
DEFAULT_TAGS = load_json_file("default_tags.json")
ROLES = load_json_file("roles.json")
HAIRSTYLES = load_json_file("hairstyles.json")
ADDITIONAL_DETAILS = load_json_file("additional_details.json")
PHOTOGRAPHY_STYLES = load_json_file("photography_styles.json")
DEVICE = load_json_file("device.json")
PHOTOGRAPHER = load_json_file("photographer.json")
ARTIST = load_json_file("artist.json")
DIGITAL_ARTFORM = load_json_file("digital_artform.json")
PLACE = load_json_file("place.json")
LIGHTING = load_json_file("lighting.json")
CLOTHING = load_json_file("clothing.json")
COMPOSITION = load_json_file("composition.json")
POSE = load_json_file("pose.json")
BACKGROUND = load_json_file("background.json")
class PromptGenerator:
def __init__(self, seed=None):
self.rng = random.Random(seed)
def split_and_choose(self, input_str):
choices = [choice.strip() for choice in input_str.split(",")]
return self.rng.choices(choices, k=1)[0]
def get_choice(self, input_str, default_choices):
if input_str.lower() == "disabled":
return ""
elif "," in input_str:
return self.split_and_choose(input_str)
elif input_str.lower() == "random":
return self.rng.choices(default_choices, k=1)[0]
else:
return input_str
def clean_consecutive_commas(self, input_string):
cleaned_string = re.sub(r',\s*,', ',', input_string)
return cleaned_string
def process_string(self, replaced, seed):
replaced = re.sub(r'\s*,\s*', ',', replaced)
replaced = re.sub(r',+', ',', replaced)
original = replaced
first_break_clipl_index = replaced.find("BREAK_CLIPL")
second_break_clipl_index = replaced.find("BREAK_CLIPL", first_break_clipl_index + len("BREAK_CLIPL"))
if first_break_clipl_index != -1 and second_break_clipl_index != -1:
clip_content_l = replaced[first_break_clipl_index + len("BREAK_CLIPL"):second_break_clipl_index]
replaced = replaced[:first_break_clipl_index].strip(", ") + replaced[second_break_clipl_index + len("BREAK_CLIPL"):].strip(", ")
clip_l = clip_content_l
else:
clip_l = ""
first_break_clipg_index = replaced.find("BREAK_CLIPG")
second_break_clipg_index = replaced.find("BREAK_CLIPG", first_break_clipg_index + len("BREAK_CLIPG"))
if first_break_clipg_index != -1 and second_break_clipg_index != -1:
clip_content_g = replaced[first_break_clipg_index + len("BREAK_CLIPG"):second_break_clipg_index]
replaced = replaced[:first_break_clipg_index].strip(", ") + replaced[second_break_clipg_index + len("BREAK_CLIPG"):].strip(", ")
clip_g = clip_content_g
else:
clip_g = ""
t5xxl = replaced
original = original.replace("BREAK_CLIPL", "").replace("BREAK_CLIPG", "")
original = re.sub(r'\s*,\s*', ',', original)
original = re.sub(r',+', ',', original)
clip_l = re.sub(r'\s*,\s*', ',', clip_l)
clip_l = re.sub(r',+', ',', clip_l)
clip_g = re.sub(r'\s*,\s*', ',', clip_g)
clip_g = re.sub(r',+', ',', clip_g)
if clip_l.startswith(","):
clip_l = clip_l[1:]
if clip_g.startswith(","):
clip_g = clip_g[1:]
if original.startswith(","):
original = original[1:]
if t5xxl.startswith(","):
t5xxl = t5xxl[1:]
return original, seed, t5xxl, clip_l, clip_g
def generate_prompt(self, seed, custom, subject, artform, photo_type, body_types, default_tags, roles, hairstyles,
additional_details, photography_styles, device, photographer, artist, digital_artform,
place, lighting, clothing, composition, pose, background, input_image):
kwargs = locals()
del kwargs['self']
seed = kwargs.get("seed", 0)
if seed is not None:
self.rng = random.Random(seed)
components = []
custom = kwargs.get("custom", "")
if custom:
components.append(custom)
is_photographer = kwargs.get("artform", "").lower() == "photography" or (
kwargs.get("artform", "").lower() == "random"
and self.rng.choice([True, False])
)
subject = kwargs.get("subject", "")
if is_photographer:
selected_photo_style = self.get_choice(kwargs.get("photography_styles", ""), PHOTOGRAPHY_STYLES)
if not selected_photo_style:
selected_photo_style = "photography"
components.append(selected_photo_style)
if kwargs.get("photography_style", "") != "disabled" and kwargs.get("default_tags", "") != "disabled" or subject != "":
components.append(" of")
default_tags = kwargs.get("default_tags", "random")
body_type = kwargs.get("body_types", "")
if not subject:
if default_tags == "random":
if body_type != "disabled" and body_type != "random":
selected_subject = self.get_choice(kwargs.get("default_tags", ""), DEFAULT_TAGS).replace("a ", "").replace("an ", "")
components.append("a ")
components.append(body_type)
components.append(selected_subject)
elif body_type == "disabled":
selected_subject = self.get_choice(kwargs.get("default_tags", ""), DEFAULT_TAGS)
components.append(selected_subject)
else:
body_type = self.get_choice(body_type, BODY_TYPES)
components.append("a ")
components.append(body_type)
selected_subject = self.get_choice(kwargs.get("default_tags", ""), DEFAULT_TAGS).replace("a ", "").replace("an ", "")
components.append(selected_subject)
elif default_tags == "disabled":
pass
else:
components.append(default_tags)
else:
if body_type != "disabled" and body_type != "random":
components.append("a ")
components.append(body_type)
elif body_type == "disabled":
pass
else:
body_type = self.get_choice(body_type, BODY_TYPES)
components.append("a ")
components.append(body_type)
components.append(subject)
params = [
("roles", ROLES),
("hairstyles", HAIRSTYLES),
("additional_details", ADDITIONAL_DETAILS),
]
for param in params:
components.append(self.get_choice(kwargs.get(param[0], ""), param[1]))
for i in reversed(range(len(components))):
if components[i] in PLACE:
components[i] += ","
break
if kwargs.get("clothing", "") != "disabled" and kwargs.get("clothing", "") != "random":
components.append(", dressed in ")
clothing = kwargs.get("clothing", "")
components.append(clothing)
elif kwargs.get("clothing", "") == "random":
components.append(", dressed in ")
clothing = self.get_choice(kwargs.get("clothing", ""), CLOTHING)
components.append(clothing)
if kwargs.get("composition", "") != "disabled" and kwargs.get("composition", "") != "random":
components.append(",")
composition = kwargs.get("composition", "")
components.append(composition)
elif kwargs.get("composition", "") == "random":
components.append(",")
composition = self.get_choice(kwargs.get("composition", ""), COMPOSITION)
components.append(composition)
if kwargs.get("pose", "") != "disabled" and kwargs.get("pose", "") != "random":
components.append(",")
pose = kwargs.get("pose", "")
components.append(pose)
elif kwargs.get("pose", "") == "random":
components.append(",")
pose = self.get_choice(kwargs.get("pose", ""), POSE)
components.append(pose)
components.append("BREAK_CLIPG")
if kwargs.get("background", "") != "disabled" and kwargs.get("background", "") != "random":
components.append(",")
background = kwargs.get("background", "")
components.append(background)
elif kwargs.get("background", "") == "random":
components.append(",")
background = self.get_choice(kwargs.get("background", ""), BACKGROUND)
components.append(background)
if kwargs.get("place", "") != "disabled" and kwargs.get("place", "") != "random":
components.append(",")
place = kwargs.get("place", "")
components.append(place)
elif kwargs.get("place", "") == "random":
components.append(",")
place = self.get_choice(kwargs.get("place", ""), PLACE)
components.append(place + ",")
lighting = kwargs.get("lighting", "").lower()
if lighting == "random":
selected_lighting = ", ".join(self.rng.sample(LIGHTING, self.rng.randint(2, 5)))
components.append(",")
components.append(selected_lighting)
elif lighting == "disabled":
pass
else:
components.append(", ")
components.append(lighting)
components.append("BREAK_CLIPG")
components.append("BREAK_CLIPL")
if is_photographer:
if kwargs.get("photo_type", "") != "disabled":
photo_type_choice = self.get_choice(kwargs.get("photo_type", ""), PHOTO_TYPE)
if photo_type_choice and photo_type_choice != "random" and photo_type_choice != "disabled":
random_value = round(self.rng.uniform(1.1, 1.5), 1)
components.append(f", ({photo_type_choice}:{random_value}), ")
params = [
("device", DEVICE),
("photographer", PHOTOGRAPHER),
]
components.extend([self.get_choice(kwargs.get(param[0], ""), param[1]) for param in params])
if kwargs.get("device", "") != "disabled":
components[-2] = f", shot on {components[-2]}"
if kwargs.get("photographer", "") != "disabled":
components[-1] = f", photo by {components[-1]}"
else:
digital_artform_choice = self.get_choice(kwargs.get("digital_artform", ""), DIGITAL_ARTFORM)
if digital_artform_choice:
components.append(f"{digital_artform_choice}")
if kwargs.get("artist", "") != "disabled":
components.append(f"by {self.get_choice(kwargs.get('artist', ''), ARTIST)}")
components.append("BREAK_CLIPL")
prompt = " ".join(components)
prompt = re.sub(" +", " ", prompt)
replaced = prompt.replace("of as", "of")
replaced = self.clean_consecutive_commas(replaced)
return self.process_string(replaced, seed)
def add_caption_to_prompt(self, prompt, caption):
if caption:
return f"{prompt}, {caption}"
return prompt
class HuggingFaceInferenceNode:
def __init__(self):
self.clients = {
"Mixtral": InferenceClient("NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO"),
"Mistral": InferenceClient("mistralai/Mistral-7B-Instruct-v0.3"),
"Llama 3": InferenceClient("meta-llama/Meta-Llama-3-8B-Instruct"),
"Mistral-Nemo": InferenceClient("mistralai/Mistral-Nemo-Instruct-2407")
}
self.prompts_dir = "./prompts"
os.makedirs(self.prompts_dir, exist_ok=True)
def save_prompt(self, prompt):
filename_text = "hf_" + prompt.split(',')[0].strip()
filename_text = re.sub(r'[^\w\-_\. ]', '_', filename_text)
filename_text = filename_text[:30]
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
base_filename = f"{filename_text}_{timestamp}.txt"
filename = os.path.join(self.prompts_dir, base_filename)
with open(filename, "w") as file:
file.write(prompt)
print(f"Prompt saved to {filename}")
def generate(self, model, input_text, happy_talk, compress, compression_level, poster, custom_base_prompt=""):
try:
client = self.clients[model]
default_happy_prompt = """Create a detailed visually descriptive caption of this description, which will be used as a prompt for a text to image AI system (caption only, no instructions like "create an image").Remove any mention of digital artwork or artwork style. Give detailed visual descriptions of the character(s), including ethnicity, skin tone, expression etc. Imagine using keywords for a still for someone who has aphantasia. Describe the image style, e.g. any photographic or art styles / techniques utilized. Make sure to fully describe all aspects of the cinematography, with abundant technical details and visual descriptions. If there is more than one image, combine the elements and characters from all of the images creatively into a single cohesive composition with a single background, inventing an interaction between the characters. Be creative in combining the characters into a single cohesive scene. Focus on two primary characters (or one) and describe an interesting interaction between them, such as a hug, a kiss, a fight, giving an object, an emotional reaction / interaction. If there is more than one background in the images, pick the most appropriate one. Your output is only the caption itself, no comments or extra formatting. The caption is in a single long paragraph. If you feel the images are inappropriate, invent a new scene / characters inspired by these. Additionally, incorporate a specific movie director's visual style (e.g. Wes Anderson, Christopher Nolan, Quentin Tarantino) and describe the lighting setup in detail, including the type, color, and placement of light sources to create the desired mood and atmosphere. Always frame the scene as a screen grab from a 35mm film still, including details about the film grain, color grading, and any artifacts or characteristics specific to 35mm film photography."""
default_simple_prompt = """Create a brief, straightforward caption for this description, suitable for a text-to-image AI system. Focus on the main elements, key characters, and overall scene without elaborate details. Provide a clear and concise description in one or two sentences."""
poster_prompt = """Analyze the provided description and extract key information to create a movie poster style description. Format the output as follows:
Title: A catchy, intriguing title that captures the essence of the scene, place the title in "".
Main character: Give a description of the main character.
Background: Describe the background in detail.
Supporting characters: Describe the supporting characters
Branding type: Describe the branding type
Tagline: Include a tagline that captures the essence of the movie.
Visual style: Ensure that the visual style fits the branding type and tagline.
You are allowed to make up film and branding names, and do them like 80's, 90's or modern movie posters."""
if poster:
base_prompt = poster_prompt
elif custom_base_prompt.strip():
base_prompt = custom_base_prompt
else:
base_prompt = default_happy_prompt if happy_talk else default_simple_prompt
if compress and not poster:
compression_chars = {
"soft": 600 if happy_talk else 300,
"medium": 400 if happy_talk else 200,
"hard": 200 if happy_talk else 100
}
char_limit = compression_chars[compression_level]
base_prompt += f" Compress the output to be concise while retaining key visual details. MAX OUTPUT SIZE no more than {char_limit} characters."
messages = f"<|im_start|>system\nYou are a helpful assistant. Try your best to give best response possible to user.<|im_end|>"
messages += f"\n<|im_start|>user\n{base_prompt}\nDescription: {input_text}<|im_end|>\n<|im_start|>assistant\n"
stream = client.text_generation(messages, max_new_tokens=4000, do_sample=True, stream=True, details=True, return_full_text=False)
output = ""
for response in stream:
if not response.token.text == "<|im_end|>":
output += response.token.text
self.save_prompt(output)
return output
except Exception as e:
print(f"An error occurred: {e}")
return f"Error occurred while processing the request: {str(e)}"
def create_interface():
prompt_generator = PromptGenerator()
huggingface_node = HuggingFaceInferenceNode()
with gr.Blocks(theme='bethecloud/storj_theme') as demo:
gr.Markdown("# AI Prompt Generator and Text Generator")
with gr.Row():
with gr.Column(scale=2):
with gr.Accordion("Basic Settings"):
seed = gr.Number(label="Seed", value=0)
custom = gr.Textbox(label="Custom")
subject = gr.Textbox(label="Subject")
artform = gr.Dropdown(["disabled"] + ARTFORM, label="Artform", value="disabled")
photo_type = gr.Dropdown(["disabled"] + PHOTO_TYPE, label="Photo Type", value="disabled")
with gr.Accordion("Character Details", value=False):
body_types = gr.Dropdown(["disabled"] + BODY_TYPES, label="Body Types", value="disabled")
default_tags = gr.Dropdown(["disabled"] + DEFAULT_TAGS, label="Default Tags", value="disabled")
roles = gr.Dropdown(["disabled"] + ROLES, label="Roles", value="disabled")
hairstyles = gr.Dropdown(["disabled"] + HAIRSTYLES, label="Hairstyles", value="disabled")
clothing = gr.Dropdown(["disabled"] + CLOTHING, label="Clothing", value="disabled")
with gr.Accordion("Scene Details", value=False):
place = gr.Dropdown(["disabled"] + PLACE, label="Place", value="disabled")
lighting = gr.Dropdown(["disabled"] + LIGHTING, label="Lighting", value="disabled")
composition = gr.Dropdown(["disabled"] + COMPOSITION, label="Composition", value="disabled")
pose = gr.Dropdown(["disabled"] + POSE, label="Pose", value="disabled")
background = gr.Dropdown(["disabled"] + BACKGROUND, label="Background", value="disabled")
with gr.Accordion("Style and Artist", value=False):
additional_details = gr.Dropdown(["disabled"] + ADDITIONAL_DETAILS, label="Additional Details", value="disabled")
photography_styles = gr.Dropdown(["disabled"] + PHOTOGRAPHY_STYLES, label="Photography Styles", value="disabled")
device = gr.Dropdown(["disabled"] + DEVICE, label="Device", value="disabled")
photographer = gr.Dropdown(["disabled"] + PHOTOGRAPHER, label="Photographer", value="disabled")
artist = gr.Dropdown(["disabled"] + ARTIST, label="Artist", value="disabled")
digital_artform = gr.Dropdown(["disabled"] + DIGITAL_ARTFORM, label="Digital Artform", value="disabled")
with gr.Column(scale=2):
with gr.Accordion("Image and Caption", open=True):
input_image = gr.Image(label="Input Image (optional)")
caption_output = gr.Textbox(label="Generated Caption", lines=3)
create_caption_button = gr.Button("Create Caption")
with gr.Accordion("Prompt Generation", open=True):
generate_button = gr.Button("Generate Prompt")
output = gr.Textbox(label="Generated Prompt / Input Text", lines=5)
add_caption_button = gr.Button("Add Caption to Prompt")
t5xxl_output = gr.Textbox(label="T5XXL Output", visible=True)
clip_l_output = gr.Textbox(label="CLIP L Output", visible=True)
clip_g_output = gr.Textbox(label="CLIP G Output", visible=True)
with gr.Column(scale=2):
with gr.Accordion("Text Generation", open=True):
model = gr.Dropdown(["Mixtral", "Mistral", "Llama 3", "Mistral-Nemo"], label="Model", value="Mixtral")
happy_talk = gr.Checkbox(label="Happy Talk", value=True)
compress = gr.Checkbox(label="Compress", value=False)
compression_level = gr.Radio(["soft", "medium", "hard"], label="Compression Level", value="medium")
poster = gr.Checkbox(label="Poster", value=False)
custom_base_prompt = gr.Textbox(label="Custom Base Prompt", lines=5)
generate_text_button = gr.Button("Generate Text")
text_output = gr.Textbox(label="Generated Text", lines=10)
def create_caption(image):
if image is not None:
return florence_caption(image)
return ""
create_caption_button.click(
create_caption,
inputs=[input_image],
outputs=[caption_output]
)
generate_button.click(
prompt_generator.generate_prompt,
inputs=[seed, custom, subject, artform, photo_type, body_types, default_tags, roles, hairstyles,
additional_details, photography_styles, device, photographer, artist, digital_artform,
place, lighting, clothing, composition, pose, background],
outputs=[output, gr.Number(visible=False), t5xxl_output, clip_l_output, clip_g_output]
)
add_caption_button.click(
prompt_generator.add_caption_to_prompt,
inputs=[output, caption_output],
outputs=[output]
)
generate_text_button.click(
huggingface_node.generate,
inputs=[model, output, happy_talk, compress, compression_level, poster, custom_base_prompt],
outputs=text_output
)
return demo
if __name__ == "__main__":
demo = create_interface()
demo.launch() |