Spaces:
Sleeping
Sleeping
File size: 8,176 Bytes
f65c76f 29f02ac f65c76f 29f02ac f65c76f 29f02ac f65c76f 29f02ac f65c76f 29f02ac f65c76f 29f02ac f65c76f 29f02ac f65c76f 29f02ac f65c76f 29f02ac f65c76f 29f02ac f65c76f |
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 |
import logging
from pathlib import Path
import matplotlib.pyplot as plt
import torch
from diffusers import StableDiffusionPipeline
from fastcore.all import concat
from huggingface_hub import notebook_login
from PIL import Image
import numpy as np
# from IPython.display import display
from torchvision import transforms as tfms
from transformers import CLIPTextModel, CLIPTokenizer
from diffusers import AutoencoderKL, UNet2DConditionModel
from diffusers import LMSDiscreteScheduler
from tqdm.auto import tqdm
logging.disable(logging.WARNING)
class ImageGenerator():
def __init__(self,
g:int=7.5,
):
self.latent_images = []
self.g = g
self.width = 512
self.height = 512
self.generator = torch.manual_seed(32)
self.bs = 1
def __repr__(self):
return f"Image Generator with {self.g=}"
def load_models(self):
self.tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14", torch_dtype=torch.float16)
self.text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14", torch_dtype=torch.float16 ).to("cuda")
# vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-ema", torch_dtype=torch.float16 ).to("cuda")
self.vae = AutoencoderKL.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="vae" ).to("cuda")
self.unet = UNet2DConditionModel.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="unet" ).to("cuda") #torch_dtype=torch.float16,
def load_scheduler( self,
beta_start : float=0.00085,
beta_end : float=0.012,
beta_schedule : str="scaled_linear",
num_train_timesteps :int=1000):
self.scheduler = LMSDiscreteScheduler(
beta_start=beta_start,
beta_end=beta_end,
beta_schedule="scaled_linear",
num_train_timesteps=num_train_timesteps)
def load_image(self, filepath:str):
return Image.open(filepath).resize(size=(self.width,self.height))
#.convert("RGB") # RGB = 3 dimensions, RGBA = 4 dimensions
def pil_to_latent(self, image: Image) -> torch.Tensor:
with torch.no_grad():
np_img = np.transpose( (( np.array(image) / 255)-0.5)*2, (2,0,1)) # turn pil image into np array with values between -1 and 1
# print(f"{np_img.shape=}") # 4, 64, 64
np_images = np.repeat(np_img[np.newaxis, :, :], self.bs, axis=0) # adding a new dimension and repeating the image for each prompt
# print(f"{np_images.shape=}")
decoded_latent = torch.from_numpy(np_images).to("cuda").float() #<-- stability-ai vae uses half(), compvis vae uses float?
# print(f"{decoded_latent.shape=}")
encoded_latent = 0.18215 * self.vae.encode(decoded_latent).latent_dist.sample()
# print(f"{encoded_latent.shape=}")
return encoded_latent
def add_noise(self, latent: torch.Tensor, scheduler_steps: int = 10) -> torch.FloatTensor:
# noise = torch.randn_like(latent) # missing generator parameter
noise = torch.randn(
size = (self.bs, self.unet.config.in_channels, self.height//8, self.width//8),
generator = self.generator).to("cuda")
timesteps = torch.tensor([self.scheduler.timesteps[scheduler_steps]])
noisy_latent = self.scheduler.add_noise(latent, noise, timesteps)
# print(f"add_noise: {timesteps.shape=} {timesteps=} {noisy_latent.shape=}")
return noisy_latent
def latent_to_pil(self, latent:torch.Tensor) -> Image:
# print(f"latent_to_pil {latent.dtype=}")
with torch.no_grad():
decoded = self.vae.decode(1 / 0.18215 * latent).sample[0]
# print(f"latent_to_pil {decoded.shape=}")
image = (decoded/2+0.5).clamp(0,1).detach().cpu().permute(1, 2, 0).numpy()
return Image.fromarray((image*255).round().astype("uint8"))
def image_grid(self, imgs: [Image]) -> Image:
w,h = imgs[0].size
cols = len(imgs)
grid = Image.new('RGB', size=(cols*w, h))
for i, img in enumerate(imgs):
# print(f"{img.size=}")
grid.paste(img, box=(i%cols*w, i//cols*h))
return grid
def text_enc(self, prompt:str, maxlen=None) -> torch.Tensor:
'''tokenize and encode a prompt'''
if maxlen is None: maxlen = self.tokenizer.model_max_length
inp = self.tokenizer([prompt], padding="max_length", max_length=maxlen, truncation=True, return_tensors="pt")
return self.text_encoder(inp.input_ids.to("cuda"))[0].float()
def tensor_to_pil(self, t:torch.Tensor) -> Image:
'''transforms a tensor decoded by the vae to a pil image'''
# print(f"tensor_to_pil {t.shape=} {type(t)=}")
image = (t/2+0.5).clamp(0,1).detach().cpu().permute(1, 2, 0).numpy()
return Image.fromarray((image*255).round().astype("uint8"))
def latent_callback(self, latent:torch.Tensor) -> None:
'''store latents in an array so that we can inpect them later.'''
with torch.no_grad():
# print(f"cb {latent.shape=}")
decoded = self.vae.decode(1 / 0.18215 * latent).sample[0]
self.latent_images.append(self.tensor_to_pil(decoded))
def generate(self,
prompt : str,
secondary_prompt: str=None,
prompt_mix_ratio : float=0.5,
negative_prompt="",
seed : int=32,
steps : int=30,
start_step_ratio : float=1/5,
init_image : str=None,
latent_callback_mod : int=10):
self.latent_images = []
if not negative_prompt: negative_prompt = ""
with torch.no_grad():
text = self.text_enc(prompt)
if secondary_prompt:
sec_prompt_text = self.text_enc(secondary_prompt)
text = text * prompt_mix_ratio + sec_prompt_text * ( 1 - prompt_mix_ratio )
uncond = self.text_enc(negative_prompt * self.bs, text.shape[1])
emb = torch.cat([uncond, text])
if seed: torch.manual_seed(seed)
self.scheduler.set_timesteps(steps)
self.scheduler.timesteps = self.scheduler.timesteps.to(torch.float32)
if (init_image == None):
start_steps = 0
latents = torch.randn(
size = (self.bs, self.unet.config.in_channels, self.height//8, self.width//8),
generator = self.generator)
latents = latents * self.scheduler.init_noise_sigma
# print(f"{latents.shape=}")
else:
start_steps = int(steps * start_step_ratio) # 0%: too much noise, 100% no noise
# print(f"{start_steps=}")
img = self.load_image(init_image)
latents =self. pil_to_latent(img)
self.latent_callback(latents)
latents = self.add_noise(latents, start_steps).to("cuda").float()
self.latent_callback(latents)
latents = latents.to("cuda").float()
for i,ts in enumerate(tqdm(self.scheduler.timesteps, leave=False)):
if i >= start_steps:
inp = self.scheduler.scale_model_input(torch.cat([latents] * 2), ts)
with torch.no_grad():
u,t = self.unet(inp, ts, encoder_hidden_states=emb).sample.chunk(2) #todo, grab those with callbacks
pred = u + self.g*(t-u)
# pred = u + self.g*(t-u)/torch.norm(t-u)*torch.norm(u)
latents = self.scheduler.step(pred, ts, latents).prev_sample
if latent_callback_mod and i % latent_callback_mod == 0:
self.latent_callback(latents)
return self.latent_to_pil(latents), self.latent_images |