146 lines
5.2 KiB
Python
146 lines
5.2 KiB
Python
from hashlib import blake2b
|
||
import argon2
|
||
|
||
import base64
|
||
import io
|
||
import re
|
||
import requests
|
||
from requests.adapters import HTTPAdapter, Retry
|
||
import comfy.utils
|
||
|
||
import torch
|
||
import numpy as np
|
||
from PIL import Image, ImageOps
|
||
|
||
# cherry-picked from novelai_api.utils
|
||
def argon_hash(email: str, password: str, size: int, domain: str) -> str:
|
||
pre_salt = f"{password[:6]}{email}{domain}"
|
||
blake = blake2b(digest_size=16)
|
||
blake.update(pre_salt.encode())
|
||
salt = blake.digest()
|
||
raw = argon2.low_level.hash_secret_raw(password.encode(), salt, 2, int(2000000 / 1024), 1, size, argon2.low_level.Type.ID,)
|
||
hashed = base64.urlsafe_b64encode(raw).decode()
|
||
return hashed
|
||
|
||
def get_access_key(email: str, password: str) -> str:
|
||
return argon_hash(email, password, 64, "novelai_data_access_key")[:64]
|
||
|
||
|
||
def login(key) -> str:
|
||
response = requests.post(f"https://api.novelai.net/user/login", json={ "key": key })
|
||
response.raise_for_status()
|
||
return response.json()["accessToken"]
|
||
|
||
BASE_URL="https://image.novelai.net"
|
||
def generate_image(access_token, prompt, model, action, parameters, timeout=None, retry=None):
|
||
data = { "input": prompt, "model": model, "action": action, "parameters": parameters }
|
||
|
||
request = requests
|
||
if retry is not None and retry > 1:
|
||
retries = Retry(total=retry, backoff_factor=1, status_forcelist=[429, 500, 502, 503, 504], allowed_methods=["POST"])
|
||
session = requests.Session()
|
||
session.mount("https://", HTTPAdapter(max_retries=retries))
|
||
request = session
|
||
|
||
response = request.post(f"{BASE_URL}/ai/generate-image", json=data, headers={ "Authorization": f"Bearer {access_token}" }, timeout=timeout)
|
||
response.raise_for_status()
|
||
return response.content
|
||
|
||
|
||
def image_to_base64(image):
|
||
i = 255. * image[0].cpu().numpy()
|
||
img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8))
|
||
image_bytesIO = io.BytesIO()
|
||
img.save(image_bytesIO, format="png")
|
||
return base64.b64encode(image_bytesIO.getvalue()).decode()
|
||
|
||
def naimask_to_base64(image):
|
||
i = 255. * image[0].cpu().numpy()
|
||
i = np.clip(i, 0, 255).astype(np.uint8)
|
||
alpha = np.sum(i, axis=-1) > 0
|
||
alpha = np.uint8(alpha * 255)
|
||
rgba = np.dstack((i, alpha))
|
||
img = Image.fromarray(rgba)
|
||
image_bytesIO = io.BytesIO()
|
||
img.save(image_bytesIO, format="png")
|
||
return base64.b64encode(image_bytesIO.getvalue()).decode()
|
||
|
||
def bytes_to_image(image_bytes):
|
||
i = Image.open(io.BytesIO(image_bytes))
|
||
i = ImageOps.exif_transpose(i)
|
||
image = np.array(i).astype(np.float32) / 255.0
|
||
return torch.from_numpy(image)[None,]
|
||
|
||
def blank_image():
|
||
return torch.tensor([[[0]]])
|
||
|
||
def resize_image(image, size_to):
|
||
samples = image.movedim(-1,1)
|
||
w, h = size_to
|
||
s = comfy.utils.common_upscale(samples, w, h, "bilinear", "disabled")
|
||
s = s.movedim(1,-1)
|
||
return s
|
||
|
||
def resize_to_naimask(mask, image_size=None):
|
||
samples = mask.movedim(-1,1)
|
||
w, h = (samples.shape[3], samples.shape[2]) if not image_size else image_size
|
||
width = int(np.ceil(w / 64) * 8)
|
||
height = int(np.ceil(h / 64) * 8)
|
||
s = comfy.utils.common_upscale(samples, width, height, "nearest-exact", "disabled")
|
||
s = s.movedim(1,-1)
|
||
return s
|
||
|
||
def calculate_resolution(pixel_count, aspect_ratio):
|
||
pixel_count = pixel_count / 4096
|
||
w, h = aspect_ratio
|
||
k = (pixel_count * w / h) ** 0.5
|
||
width = int(np.floor(k) * 64)
|
||
height = int(np.floor(k * h / w) * 64)
|
||
return width, height
|
||
|
||
|
||
def prompt_to_stack(sentence):
|
||
result = []
|
||
current_str = ""
|
||
stack = [{ "weight": 1.0, "data": result }]
|
||
|
||
for i, c in enumerate(sentence):
|
||
if c in '()':
|
||
# current_str = current_str.strip()
|
||
if c == '(':
|
||
if current_str: stack[-1]["data"].append(current_str)
|
||
stack[-1]["data"].append({ "weight": 1.0, "data": [] });
|
||
stack.append(stack[-1]["data"][-1])
|
||
elif c == ')':
|
||
searched = re.search(r"^(.*):([0-9\.]+)$", current_str)
|
||
current_str, weight = searched.groups() if searched else (current_str, 1.1)
|
||
if current_str: stack[-1]["data"].append(current_str)
|
||
stack[-1]["weight"] = float(weight)
|
||
if stack[-1]["data"] != result:
|
||
stack.pop()
|
||
else: # no more to pop
|
||
print("error :", sentence);
|
||
print(f"col {i:>3}:", " " * i + "^")
|
||
# raise Exception('Error durring parsing parentheses', sentence, i, c)
|
||
current_str = ""
|
||
else:
|
||
current_str += c
|
||
|
||
if current_str:
|
||
stack[-1]["data"].append(current_str)
|
||
|
||
return result
|
||
|
||
def prompt_stack_to_nai(l, weight_per_brace=0.05):
|
||
result = ""
|
||
for el in l:
|
||
if isinstance(el, dict):
|
||
brace_count = round((el["weight"] - 1.0) / weight_per_brace)
|
||
result += "{" * brace_count + "[" * -brace_count + prompt_stack_to_nai(el["data"]) + "}" * brace_count + "]" * -brace_count
|
||
else:
|
||
result += el
|
||
return result
|
||
|
||
def prompt_to_nai(prompt, weight_per_brace=0.05):
|
||
return prompt_stack_to_nai(prompt_to_stack(prompt.replace("\(", "（").replace("\)", "）")), weight_per_brace).replace("（", "(").replace("）",")")
|