File size: 5,187 Bytes
c978a88 c8e1bf7 e186a0a c8e1bf7 e186a0a c8e1bf7 e186a0a c8e1bf7 e186a0a c8e1bf7 e186a0a c8e1bf7 |
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 |
from typing import Union, Callable, Optional
import torch
import PIL
from diffusers import (
AutoencoderKL,
DDIMScheduler,
DiffusionPipeline,
LMSDiscreteScheduler,
PNDMScheduler,
UNet2DConditionModel,
)
#from PIL import Image
from torchvision import transforms as tfms
from tqdm.auto import tqdm
from transformers import CLIPTextModel, CLIPTokenizer
class MagicMixPipeline(DiffusionPipeline):
def __init__(
self,
vae: AutoencoderKL,
text_encoder: CLIPTextModel,
tokenizer: CLIPTokenizer,
unet: UNet2DConditionModel,
scheduler: Union[PNDMScheduler, LMSDiscreteScheduler, DDIMScheduler],
):
super().__init__()
self.register_modules(vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler)
# convert PIL image to latents
def encode(self, img):
with torch.no_grad():
latent = self.vae.encode(tfms.ToTensor()(img).unsqueeze(0).to(self.device) * 2 - 1)
latent = 0.18215 * latent.latent_dist.sample()
return latent
# convert latents to PIL image
def decode(self, latent):
latent = (1 / 0.18215) * latent
with torch.no_grad():
img = self.vae.decode(latent).sample
img = (img / 2 + 0.5).clamp(0, 1)
img = img.detach().cpu().permute(0, 2, 3, 1).numpy()
img = (img * 255).round().astype("uint8")
return PIL.Image.fromarray(img[0])
# convert prompt into text embeddings, also unconditional embeddings
def prep_text(self, prompt):
text_input = self.tokenizer(
prompt,
padding="max_length",
max_length=self.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_embedding = self.text_encoder(text_input.input_ids.to(self.device))[0]
uncond_input = self.tokenizer(
"",
padding="max_length",
max_length=self.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
uncond_embedding = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
return torch.cat([uncond_embedding, text_embedding])
def __call__(
self,
img: PIL.Image.Image,
prompt: str,
kmin: float = 0.3,
kmax: float = 0.6,
mix_factor: float = 0.5,
seed: int = 42,
steps: int = 50,
guidance_scale: float = 7.5,
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
callback_steps: Optional[int] = 1,
) -> PIL.Image.Image:
tmin = steps - int(kmin * steps)
tmax = steps - int(kmax * steps)
text_embeddings = self.prep_text(prompt)
self.scheduler.set_timesteps(steps)
width, height = img.size
encoded = self.encode(img)
torch.manual_seed(seed)
noise = torch.randn(
(1, self.unet.in_channels, height // 8, width // 8),
).to(self.device)
latents = self.scheduler.add_noise(
encoded,
noise,
timesteps=self.scheduler.timesteps[tmax],
)
input = torch.cat([latents] * 2)
input = self.scheduler.scale_model_input(input, self.scheduler.timesteps[tmax])
with torch.no_grad():
pred = self.unet(
input,
self.scheduler.timesteps[tmax],
encoder_hidden_states=text_embeddings,
).sample
pred_uncond, pred_text = pred.chunk(2)
pred = pred_uncond + guidance_scale * (pred_text - pred_uncond)
latents = self.scheduler.step(pred, self.scheduler.timesteps[tmax], latents).prev_sample
for i, t in enumerate(tqdm(self.scheduler.timesteps)):
if i > tmax:
if i < tmin: # layout generation phase
orig_latents = self.scheduler.add_noise(
encoded,
noise,
timesteps=t,
)
input = (mix_factor * latents) + (
1 - mix_factor
) * orig_latents # interpolating between layout noise and conditionally generated noise to preserve layout sematics
input = torch.cat([input] * 2)
else: # content generation phase
input = torch.cat([latents] * 2)
input = self.scheduler.scale_model_input(input, t)
with torch.no_grad():
pred = self.unet(
input,
t,
encoder_hidden_states=text_embeddings,
).sample
pred_uncond, pred_text = pred.chunk(2)
pred = pred_uncond + guidance_scale * (pred_text - pred_uncond)
latents = self.scheduler.step(pred, t, latents).prev_sample
# call the callback, if provided
if callback is not None and i % callback_steps == 0:
callback(i, t, latents)
return self.decode(latents)
|