Create super-res.py
Browse files- super-res.py +176 -0
super-res.py
ADDED
@@ -0,0 +1,176 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import sys, argparse, glob, os
|
2 |
+
import torch
|
3 |
+
import numpy as np
|
4 |
+
from tqdm import tqdm
|
5 |
+
import gradio as gr
|
6 |
+
from PIL import Image
|
7 |
+
from omegaconf import OmegaConf
|
8 |
+
from einops import repeat, rearrange
|
9 |
+
from pytorch_lightning import seed_everything
|
10 |
+
from imwatermark import WatermarkEncoder
|
11 |
+
|
12 |
+
from scripts.txt2img import put_watermark
|
13 |
+
from ldm.models.diffusion.ddim import DDIMSampler
|
14 |
+
from ldm.models.diffusion.ddpm import LatentUpscaleDiffusion, LatentUpscaleFinetuneDiffusion
|
15 |
+
from ldm.util import exists, instantiate_from_config
|
16 |
+
|
17 |
+
|
18 |
+
torch.set_grad_enabled(False)
|
19 |
+
|
20 |
+
|
21 |
+
def load_model_from_config(config, ckpt, verbose=False):
|
22 |
+
print(f"Loading model from {ckpt}")
|
23 |
+
pl_sd = torch.load(ckpt, map_location="cpu")
|
24 |
+
if "global_step" in pl_sd:
|
25 |
+
print(f"Global Step: {pl_sd['global_step']}")
|
26 |
+
sd = pl_sd["state_dict"]
|
27 |
+
model = instantiate_from_config(config.model)
|
28 |
+
m, u = model.load_state_dict(sd, strict=False)
|
29 |
+
if len(m) > 0 and verbose:
|
30 |
+
print("missing keys:")
|
31 |
+
print(m)
|
32 |
+
if len(u) > 0 and verbose:
|
33 |
+
print("unexpected keys:")
|
34 |
+
print(u)
|
35 |
+
|
36 |
+
model.cuda()
|
37 |
+
model.eval()
|
38 |
+
return model
|
39 |
+
|
40 |
+
|
41 |
+
def make_batch_sd( image, txt, device,num_samples=1,size=(512,512)):
|
42 |
+
image = Image.open(image).convert("RGB")
|
43 |
+
image = image.resize(size)
|
44 |
+
image = np.array(image)
|
45 |
+
image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
|
46 |
+
batch = {
|
47 |
+
"lr": rearrange(image, 'h w c -> 1 c h w'),
|
48 |
+
"txt": num_samples * [txt],
|
49 |
+
}
|
50 |
+
batch["lr"] = repeat(batch["lr"].to(device=device), "1 ... -> n ...", n=num_samples)
|
51 |
+
return batch
|
52 |
+
|
53 |
+
|
54 |
+
def make_noise_augmentation(model, batch, noise_level=None):
|
55 |
+
x_low = batch[model.low_scale_key]
|
56 |
+
x_low = x_low.to(memory_format=torch.contiguous_format).float()
|
57 |
+
x_aug, noise_level = model.low_scale_model(x_low, noise_level)
|
58 |
+
return x_aug, noise_level
|
59 |
+
|
60 |
+
|
61 |
+
def paint(sampler, image, prompt, seed, scale, h, w, steps, num_samples=1, callback=None, eta=0., noise_level=None):
|
62 |
+
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
63 |
+
model = sampler.model
|
64 |
+
seed_everything(seed)
|
65 |
+
prng = np.random.RandomState(seed)
|
66 |
+
start_code = prng.randn(num_samples, model.channels, h, w)
|
67 |
+
start_code = torch.from_numpy(start_code).to(device=device, dtype=torch.float32)
|
68 |
+
|
69 |
+
with torch.no_grad(), torch.autocast("cuda"):
|
70 |
+
batch = make_batch_sd(image, txt=prompt, device=device, num_samples=num_samples, size=(h, w))
|
71 |
+
c = model.cond_stage_model.encode(batch["txt"])
|
72 |
+
c_cat = list()
|
73 |
+
if isinstance(model, LatentUpscaleFinetuneDiffusion):
|
74 |
+
for ck in model.concat_keys:
|
75 |
+
cc = batch[ck]
|
76 |
+
if exists(model.reshuffle_patch_size):
|
77 |
+
assert isinstance(model.reshuffle_patch_size, int)
|
78 |
+
cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w',p1=model.reshuffle_patch_size, p2=model.reshuffle_patch_size)
|
79 |
+
c_cat.append(cc)
|
80 |
+
c_cat = torch.cat(c_cat, dim=1)
|
81 |
+
# cond
|
82 |
+
cond = {"c_concat": [c_cat], "c_crossattn": [c]}
|
83 |
+
# uncond cond
|
84 |
+
uc_cross = model.get_unconditional_conditioning(num_samples, "")
|
85 |
+
uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]}
|
86 |
+
elif isinstance(model, LatentUpscaleDiffusion):
|
87 |
+
x_augment, noise_level = make_noise_augmentation(model, batch, noise_level)
|
88 |
+
cond = {"c_concat": [x_augment], "c_crossattn": [c], "c_adm": noise_level}
|
89 |
+
# uncond cond
|
90 |
+
uc_cross = model.get_unconditional_conditioning(num_samples, "")
|
91 |
+
uc_full = {"c_concat": [x_augment], "c_crossattn": [uc_cross], "c_adm": noise_level}
|
92 |
+
else:
|
93 |
+
raise NotImplementedError()
|
94 |
+
|
95 |
+
shape = [model.channels, h, w]
|
96 |
+
samples, intermediates = sampler.sample(
|
97 |
+
steps,
|
98 |
+
num_samples,
|
99 |
+
shape,
|
100 |
+
cond,
|
101 |
+
verbose=False,
|
102 |
+
eta=eta,
|
103 |
+
unconditional_guidance_scale=scale,
|
104 |
+
unconditional_conditioning=uc_full,
|
105 |
+
x_T=start_code,
|
106 |
+
callback=callback
|
107 |
+
)
|
108 |
+
with torch.no_grad():
|
109 |
+
x_samples_ddim = model.decode_first_stage(samples)
|
110 |
+
result = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
|
111 |
+
result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255
|
112 |
+
return Image.fromarray(result.astype(np.uint8)[0])
|
113 |
+
|
114 |
+
|
115 |
+
if __name__ == "__main__":
|
116 |
+
parser = argparse.ArgumentParser()
|
117 |
+
parser.add_argument("--indir", type=str, nargs="?", help="dir containing image-mask pairs (`example.png` and `example_mask.png`)",)
|
118 |
+
parser.add_argument("--num_imgs", type=int, default=None, help="number of images to generate",)
|
119 |
+
parser.add_argument("--steps",type=int,default=50,help="number of ddim sampling steps",)
|
120 |
+
parser.add_argument("--config",type=str,default="/checkpoint/pfz/autoencoders/sd/stable-diffusion-x4-upscaler/x4-upscaling.yaml",help="path to config which constructs model",)
|
121 |
+
parser.add_argument("--ckpt",type=str,default="/checkpoint/pfz/autoencoders/sd/stable-diffusion-x4-upscaler/x4-upscaler-ema.ckpt",help="path to checkpoint of model",)
|
122 |
+
parser.add_argument("--ldm_decoder_ckpt",default=None,type=str,help="path to checkpoint of LDM decoder")
|
123 |
+
parser.add_argument("--num_samples",default=1,type=int,help="number of samples to generate")
|
124 |
+
parser.add_argument("--scale", default=10.0, type=float, help="scale")
|
125 |
+
parser.add_argument("--eta", default=0.0, type=float, help="eta")
|
126 |
+
parser.add_argument("--noise_level", default=20, type=float, help="eta")
|
127 |
+
parser.add_argument("--output_dir",type=str,default="outputs",nargs="?",help="dir to write results to",)
|
128 |
+
parser.add_argument("--height",type=int,default=512,help="height of output image",)
|
129 |
+
parser.add_argument("--width",type=int,default=512,help="width of output image",)
|
130 |
+
parser.add_argument("--seed",type=int,default=0,help="random seed",)
|
131 |
+
opt = parser.parse_args()
|
132 |
+
|
133 |
+
print(f'>>> Building LDM model with config {opt.config} and weights from {opt.ckpt}...')
|
134 |
+
config = OmegaConf.load(f"{opt.config}")
|
135 |
+
model = load_model_from_config(config, f"{opt.ckpt}")
|
136 |
+
|
137 |
+
# Parameter None for clutil sweep
|
138 |
+
print(f'reload decoder weights {opt.ldm_decoder_ckpt}...')
|
139 |
+
if opt.ldm_decoder_ckpt is not None and opt.ldm_decoder_ckpt.lower() == "none":
|
140 |
+
opt.ldm_decoder_ckpt = None
|
141 |
+
if opt.ldm_decoder_ckpt is not None:
|
142 |
+
state_dict = torch.load(opt.ldm_decoder_ckpt)['ldm_decoder']
|
143 |
+
msg = model.first_stage_model.load_state_dict(state_dict, strict=False)
|
144 |
+
print(msg)
|
145 |
+
|
146 |
+
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
147 |
+
model = model.to(device)
|
148 |
+
model.eval()
|
149 |
+
sampler = DDIMSampler(model)
|
150 |
+
|
151 |
+
os.makedirs(opt.output_dir, exist_ok=True)
|
152 |
+
|
153 |
+
images = sorted(glob.glob(os.path.join(opt.indir, "*.png"))) + sorted(glob.glob(os.path.join(opt.indir, "*.jpg"))) + sorted(glob.glob(os.path.join(opt.indir, "*.jpeg")))
|
154 |
+
images += sorted(glob.glob(os.path.join(opt.indir, "*.PNG"))) + sorted(glob.glob(os.path.join(opt.indir, "*.JPG"))) + sorted(glob.glob(os.path.join(opt.indir, "*.JPEG")))
|
155 |
+
print(f"Found {len(images)} inputs.")
|
156 |
+
|
157 |
+
counter = 0
|
158 |
+
for image in tqdm(images):
|
159 |
+
if opt.num_imgs is not None and counter >= opt.num_imgs:
|
160 |
+
break
|
161 |
+
noise_level = torch.Tensor( opt.num_samples * [opt.noise_level]).to(sampler.model.device).long()
|
162 |
+
sampler.make_schedule(opt.steps, ddim_eta=opt.eta, verbose=True)
|
163 |
+
result = paint(
|
164 |
+
sampler=sampler,
|
165 |
+
image=image,
|
166 |
+
prompt="",
|
167 |
+
seed=opt.seed,
|
168 |
+
scale=opt.scale,
|
169 |
+
h=opt.height, w=opt.width, steps=opt.steps,
|
170 |
+
num_samples=opt.num_samples,
|
171 |
+
callback=None,
|
172 |
+
noise_level=noise_level
|
173 |
+
)
|
174 |
+
outpath = os.path.join(opt.output_dir, os.path.split(image)[1]).replace('.jpg', '.png').replace('.jpeg', '.png').replace('.JPG', '.png').replace('.JPEG', '.png')
|
175 |
+
result.save(outpath)
|
176 |
+
counter += 1
|