pixel-prompt / main.py
Hatman's picture
Update main.py
68d7b9b verified
raw
history blame
13.8 kB
import random
from fastapi import FastAPI
from fastapi.staticfiles import StaticFiles
from fastapi.responses import FileResponse
from fastapi.middleware.cors import CORSMiddleware
from huggingface_hub import InferenceClient, login
from transformers import AutoTokenizer
from pydantic import BaseModel
from gradio_client import Client, file
from starlette.responses import StreamingResponse
import re
from datetime import datetime
import json
import requests
import base64
import os
import time
from PIL import Image
from io import BytesIO
import aiohttp
import asyncio
from typing import Optional
from dotenv import load_dotenv
import boto3
from groq import Groq
app = FastAPI()
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
groqClient = Groq (api_key=os.environ.get("GROQ_API_KEY"))
load_dotenv()
token = os.environ.get("HF_TOKEN")
login(token)
prompt_model = "llama-3.1-8b-instant"
magic_prompt_model = "Gustavosta/MagicPrompt-Stable-Diffusion"
options = {"use_cache": False, "wait_for_model": True}
parameters = {"return_full_text":False, "max_new_tokens":300}
headers = {"Authorization": f"Bearer {token}", "x-use-cache":"0", 'Content-Type' :'application/json'}
API_URL = f'https://api-inference.huggingface.co/models/'
perm_negative_prompt = "watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry"
cwd = os.getcwd()
pictures_directory = os.path.join(cwd, 'pictures')
last_two_models = []
class Item(BaseModel):
prompt: str
steps: int
guidance: float
modelID: str
modelLabel: str
image: Optional[str] = None
target: str
control: float
class Core(BaseModel):
itemString: str
@app.get("/core")
async def core():
if not os.path.exists(pictures_directory):
os.makedirs(pictures_directory)
async def generator():
# Start JSON array
yield '['
first = True
for filename in os.listdir(pictures_directory):
if filename.endswith('.json'):
file_path = os.path.join(pictures_directory, filename)
with open(file_path, 'r') as file:
data = json.load(file)
# For JSON formatting, ensure only the first item doesn't have a preceding comma
if first:
first = False
else:
yield ','
yield json.dumps({"base64": data["base64image"], "prompt": data["returnedPrompt"]})
# End JSON array
yield ']'
return StreamingResponse(generator(), media_type="application/json")
def getPrompt(prompt, modelID, attempts=1):
response = {}
print(modelID)
try:
if modelID != magic_prompt_model:
chat = [
{"role": "user", "content": prompt_base},
{"role": "assistant", "content": prompt_assistant},
{"role": "user", "content": prompt},
]
response = groqClient.chat.completions.create(messages=chat, temperature=1, max_tokens=2048, top_p=1, stream=False, stop=None, model=modelID)
else:
apiData={"inputs":prompt, "parameters": parameters, "options": options, "timeout": 45}
response = requests.post(API_URL + modelID, headers=headers, data=json.dumps(apiData))
return response.json()
except Exception as e:
print(f"An error occurred: {e}")
if attempts < 3:
getPrompt(prompt, modelID, attempts + 1)
return response
@app.post("/inferencePrompt")
def inferencePrompt(item: Core):
print("Start API Inference Prompt")
try:
plain_response_data = getPrompt(item.itemString, prompt_model)
magic_response_data = getPrompt(item.itemString, magic_prompt_model)
returnJson = {"plain": plain_response_data.choices[0].message.content, "magic": item.itemString + magic_response_data[0]["generated_text"]}
print(f'Return Json {returnJson}')
return returnJson
except Exception as e:
returnJson = {"plain": f'An Error occured: {e}', "magic": f'An Error occured: {e}'}
async def wake_model(modelID):
data = {"inputs":"wake up call", "options":options}
headers = {"Authorization": f"Bearer {token}"}
api_data = json.dumps(data)
try:
timeout = aiohttp.ClientTimeout(total=60) # Set timeout to 60 seconds
async with aiohttp.ClientSession(timeout=timeout) as session:
async with session.post(API_URL + modelID, headers=headers, data=api_data) as response:
pass
print('Model Waking')
except Exception as e:
print(f"An error occurred: {e}")
def formatReturn(result):
img = Image.open(result)
img.save("test.png")
img_byte_arr = BytesIO()
img.save(img_byte_arr, format='PNG')
img_byte_arr = img_byte_arr.getvalue()
base64_img = base64.b64encode(img_byte_arr).decode('utf-8')
return base64_img
def save_image(base64image, item, model, NSFW):
if not NSFW:
data = {"base64image": "data:image/png;base64," + base64image, "returnedPrompt": "Model:\n" + model + "\n\nPrompt:\n" + item.prompt, "prompt": item.prompt, "steps": item.steps, "guidance": item.guidance, "control": item.control, "target": item.target}
timestamp = datetime.now().strftime("%Y%m%d%H%M%S")
file_path = os.path.join(pictures_directory, f'{timestamp}.json')
with open(file_path, 'w') as json_file:
json.dump(data, json_file)
def gradioSD3(item):
client = Client(item.modelID, hf_token=token)
result = client.predict(
prompt=item.prompt,
negative_prompt=perm_negative_prompt,
guidance_scale=item.guidance,
num_inference_steps=item.steps,
api_name="/infer"
)
return formatReturn(result[0])
def gradioAuraFlow(item):
client = Client("multimodalart/AuraFlow")
result = client.predict(
prompt=item.prompt,
negative_prompt=perm_negative_prompt,
randomize_seed=True,
guidance_scale=item.guidance,
num_inference_steps=item.steps,
api_name="/infer"
)
print(result[0])
return formatReturn(result[0]["value"])
def gradioHatmanInstantStyle(item):
client = Client("Hatman/InstantStyle")
image_stream = BytesIO(base64.b64decode(item.image.split("base64,")[1]))
image = Image.open(image_stream)
image.save("style.png")
result = client.predict(
image_pil=file("style.png"),
prompt=item.prompt,
n_prompt=perm_negative_prompt,
scale=1,
control_scale=item.control,
guidance_scale=item.guidance,
num_inference_steps=item.steps,
seed=1,
target=item.target,
api_name="/create_image"
)
return formatReturn(result)
def lambda_image(prompt, modelID):
data = {
"prompt": prompt,
"modelID": modelID
}
serialized_data = json.dumps(data)
aws_id = os.environ.get("AWS_ID")
aws_secret = os.environ.get("AWS_SECRET")
aws_region = os.environ.get("AWS_REGION")
try:
session = boto3.Session(aws_access_key_id=aws_id, aws_secret_access_key=aws_secret, region_name=aws_region)
lambda_client = session.client('lambda')
response = lambda_client.invoke(
FunctionName='pixel_prompt_lambda',
InvocationType='RequestResponse',
Payload=serialized_data
)
response_payload = response['Payload'].read()
response_data = json.loads(response_payload)
except Exception as e:
print(f"An error occurred: {e}")
return response_data['body']
def inferenceAPI(model, item, attempts = 1):
print(f'Inference model {model}')
if attempts > 5:
return 'An error occured when Processing', model
prompt = item.prompt
if "dallinmackay" in model:
prompt = "lvngvncnt, " + item.prompt
data = {"inputs":prompt, "negative_prompt": perm_negative_prompt, "options":options, "timeout": 45}
api_data = json.dumps(data)
try:
response = requests.request("POST", API_URL + model, headers=headers, data=api_data)
if response is None:
inferenceAPI(get_random_model(activeModels['text-to-image']), item, attempts+1)
print(response.content[0:200])
image_stream = BytesIO(response.content)
image = Image.open(image_stream)
image.save("response.png")
with open('response.png', 'rb') as f:
base64_img = base64.b64encode(f.read()).decode('utf-8')
return model, base64_img
except Exception as e:
print(f'Error When Processing Image: {e}')
activeModels = InferenceClient().list_deployed_models()
model = get_random_model(activeModels['text-to-image'])
pattern = r'^(.{1,30})\/(.{1,50})$'
if not re.match(pattern, model):
return "error model not valid", model
return inferenceAPI(model, item, attempts+1)
def get_random_model(models):
global last_two_models
model = None
priorities = [
"stabilityai/stable-diffusion-3.5-large-turbo",
"stabilityai/stable-diffusion-3.5-large",
"black-forest-labs",
"kandinsky-community",
"Kolors-diffusers",
"Juggernaut",
"insaneRealistic",
"MajicMIX",
"digiautogpt3",
"fluently"
]
for priority in priorities:
for i, model_name in enumerate(models):
if priority in model_name and model_name not in last_two_models:
model = models[i]
break
if model is not None:
break
if model is None:
print("Choosing randomly")
model = random.choice(models)
last_two_models.append(model)
last_two_models = last_two_models[-5:]
return model
def nsfw_check(item, attempts=1):
try:
API_URL = "https://api-inference.huggingface.co/models/Falconsai/nsfw_image_detection"
with open('response.png', 'rb') as f:
data = f.read()
response = requests.request("POST", API_URL, headers=headers, data=data)
decoded_response = response.content.decode("utf-8")
print(item.prompt)
print(decoded_response)
json_response = json.loads(decoded_response)
if "error" in json_response:
time.sleep(json_response["estimated_time"])
return nsfw_check(item, attempts+1)
scores = {item['label']: item['score'] for item in json_response}
error_msg = scores.get('nsfw', 0) > .1
return error_msg
except json.JSONDecodeError as e:
print(f'JSON Decoding Error: {e}')
return True
except Exception as e:
print(f'NSFW Check Error: {e}')
if attempts > 30:
return True
return nsfw_check(item, attempts+1)
@app.post("/api")
async def inference(item: Item):
print("Start API Inference")
activeModels = InferenceClient().list_deployed_models()
base64_img = ""
model = item.modelID
print(f'Start Model {model}')
NSFW = False
try:
if item.image:
model = "stabilityai/stable-diffusion-xl-base-1.0"
base64_img = gradioHatmanInstantStyle(item)
elif "AuraFlow" in item.modelID:
base64_img = gradioAuraFlow(item)
elif "Random" in item.modelID:
model = get_random_model(activeModels['text-to-image'])
pattern = r'^(.{1,30})\/(.{1,50})$'
if not re.match(pattern, model):
raise ValueError("Model not Valid")
model, base64_img= inferenceAPI(model, item)
elif "stable-diffusion-3" in item.modelID:
base64_img = gradioSD3(item)
elif "Voxel" in item.modelID or "pixel" in item.modelID:
prompt = item.prompt
if "Voxel" in item.modelID:
prompt = "voxel style, " + item.prompt
base64_img = lambda_image(prompt, item.modelID)
elif item.modelID not in activeModels['text-to-image']:
asyncio.create_task(wake_model(item.modelID))
return {"output": "Model Waking"}
else:
base64_img, model = inferenceAPI(item.modelID, item)
if 'error' in base64_img:
return {"output": base64_img, "model": model}
NSFW = nsfw_check(item)
save_image(base64_img, item, model, NSFW)
except Exception as e:
print(f"An error occurred: {e}")
base64_img = f"An error occurred: {e}"
return {"output": base64_img, "model": model, "NSFW": NSFW}
prompt_base = 'Instructions:\
\
1. Take the provided seed string as inspiration.\
2. Generate a prompt that is clear, vivid, and imaginative.\
3. This is a visual image so any reference to senses other than sight should be avoided.\
4. Ensure the prompt is between 90 and 100 tokens.\
5. Return only the prompt.\
Format your response as follows:\
Stable Diffusion Prompt: [Your prompt here]\
\
Remember:\
\
- The prompt should be descriptive.\
- Avoid overly complex or abstract phrases.\
- Make sure the prompt evokes strong imagery and can guide the creation of visual content.\
- Make sure the prompt is between 90 and 100 tokens.'
prompt_assistant = "I am ready to return a prompt that is between 90 and 100 tokens. What is your seed string?"
app.mount("/", StaticFiles(directory="web-build", html=True), name="build")
@app.get('/')
def homepage() -> FileResponse:
return FileResponse(path="/app/build/index.html", media_type="text/html")