Spaces:
Runtime error
Runtime error
add pix2pix turbo
Browse files
server/pipelines/pix2pix/__init__.py
ADDED
|
File without changes
|
server/pipelines/pix2pix/model.py
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# https://github.com/GaParmar/img2img-turbo/blob/main/src/model.py
|
| 2 |
+
from diffusers import DDPMScheduler
|
| 3 |
+
|
| 4 |
+
|
| 5 |
+
def make_1step_sched():
|
| 6 |
+
noise_scheduler_1step = DDPMScheduler.from_pretrained(
|
| 7 |
+
"stabilityai/sd-turbo", subfolder="scheduler"
|
| 8 |
+
)
|
| 9 |
+
noise_scheduler_1step.set_timesteps(1, device="cuda")
|
| 10 |
+
noise_scheduler_1step.alphas_cumprod = noise_scheduler_1step.alphas_cumprod.cuda()
|
| 11 |
+
return noise_scheduler_1step
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def my_vae_encoder_fwd(self, sample):
|
| 15 |
+
sample = self.conv_in(sample)
|
| 16 |
+
l_blocks = []
|
| 17 |
+
# down
|
| 18 |
+
for down_block in self.down_blocks:
|
| 19 |
+
l_blocks.append(sample)
|
| 20 |
+
sample = down_block(sample)
|
| 21 |
+
# middle
|
| 22 |
+
sample = self.mid_block(sample)
|
| 23 |
+
sample = self.conv_norm_out(sample)
|
| 24 |
+
sample = self.conv_act(sample)
|
| 25 |
+
sample = self.conv_out(sample)
|
| 26 |
+
self.current_down_blocks = l_blocks
|
| 27 |
+
return sample
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
def my_vae_decoder_fwd(self, sample, latent_embeds=None):
|
| 31 |
+
sample = self.conv_in(sample)
|
| 32 |
+
upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
|
| 33 |
+
# middle
|
| 34 |
+
sample = self.mid_block(sample, latent_embeds)
|
| 35 |
+
sample = sample.to(upscale_dtype)
|
| 36 |
+
if not self.ignore_skip:
|
| 37 |
+
skip_convs = [
|
| 38 |
+
self.skip_conv_1,
|
| 39 |
+
self.skip_conv_2,
|
| 40 |
+
self.skip_conv_3,
|
| 41 |
+
self.skip_conv_4,
|
| 42 |
+
]
|
| 43 |
+
# up
|
| 44 |
+
for idx, up_block in enumerate(self.up_blocks):
|
| 45 |
+
skip_in = skip_convs[idx](self.incoming_skip_acts[::-1][idx] * self.gamma)
|
| 46 |
+
# add skip
|
| 47 |
+
sample = sample + skip_in
|
| 48 |
+
sample = up_block(sample, latent_embeds)
|
| 49 |
+
else:
|
| 50 |
+
for idx, up_block in enumerate(self.up_blocks):
|
| 51 |
+
sample = up_block(sample, latent_embeds)
|
| 52 |
+
# post-process
|
| 53 |
+
if latent_embeds is None:
|
| 54 |
+
sample = self.conv_norm_out(sample)
|
| 55 |
+
else:
|
| 56 |
+
sample = self.conv_norm_out(sample, latent_embeds)
|
| 57 |
+
sample = self.conv_act(sample)
|
| 58 |
+
sample = self.conv_out(sample)
|
| 59 |
+
return sample
|
server/pipelines/pix2pix/pix2pix_turbo.py
ADDED
|
@@ -0,0 +1,212 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# https://github.com/GaParmar/img2img-turbo/blob/main/src/pix2pix_turbo.py
|
| 2 |
+
import os
|
| 3 |
+
import requests
|
| 4 |
+
import sys
|
| 5 |
+
import pdb
|
| 6 |
+
import copy
|
| 7 |
+
from tqdm import tqdm
|
| 8 |
+
import torch
|
| 9 |
+
from transformers import AutoTokenizer, PretrainedConfig, CLIPTextModel
|
| 10 |
+
from diffusers import AutoencoderKL, UNet2DConditionModel, DDPMScheduler
|
| 11 |
+
from diffusers.utils.peft_utils import set_weights_and_activate_adapters
|
| 12 |
+
from peft import LoraConfig
|
| 13 |
+
|
| 14 |
+
from pipelines.pix2pix.model import (
|
| 15 |
+
make_1step_sched,
|
| 16 |
+
my_vae_encoder_fwd,
|
| 17 |
+
my_vae_decoder_fwd,
|
| 18 |
+
)
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
class TwinConv(torch.nn.Module):
|
| 22 |
+
def __init__(self, convin_pretrained, convin_curr):
|
| 23 |
+
super(TwinConv, self).__init__()
|
| 24 |
+
self.conv_in_pretrained = copy.deepcopy(convin_pretrained)
|
| 25 |
+
self.conv_in_curr = copy.deepcopy(convin_curr)
|
| 26 |
+
self.r = None
|
| 27 |
+
|
| 28 |
+
def forward(self, x):
|
| 29 |
+
x1 = self.conv_in_pretrained(x).detach()
|
| 30 |
+
x2 = self.conv_in_curr(x)
|
| 31 |
+
return x1 * (1 - self.r) + x2 * (self.r)
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
class Pix2Pix_Turbo(torch.nn.Module):
|
| 35 |
+
def __init__(self, name, ckpt_folder="checkpoints"):
|
| 36 |
+
super().__init__()
|
| 37 |
+
self.tokenizer = AutoTokenizer.from_pretrained(
|
| 38 |
+
"stabilityai/sd-turbo", subfolder="tokenizer"
|
| 39 |
+
)
|
| 40 |
+
self.text_encoder = CLIPTextModel.from_pretrained(
|
| 41 |
+
"stabilityai/sd-turbo", subfolder="text_encoder"
|
| 42 |
+
).cuda()
|
| 43 |
+
self.sched = make_1step_sched()
|
| 44 |
+
|
| 45 |
+
vae = AutoencoderKL.from_pretrained("stabilityai/sd-turbo", subfolder="vae")
|
| 46 |
+
unet = UNet2DConditionModel.from_pretrained(
|
| 47 |
+
"stabilityai/sd-turbo", subfolder="unet"
|
| 48 |
+
)
|
| 49 |
+
|
| 50 |
+
if name == "edge_to_image":
|
| 51 |
+
url = "https://www.cs.cmu.edu/~img2img-turbo/models/edge_to_image_loras.pkl"
|
| 52 |
+
os.makedirs(ckpt_folder, exist_ok=True)
|
| 53 |
+
outf = os.path.join(ckpt_folder, "edge_to_image_loras.pkl")
|
| 54 |
+
if not os.path.exists(outf):
|
| 55 |
+
print(f"Downloading checkpoint to {outf}")
|
| 56 |
+
response = requests.get(url, stream=True)
|
| 57 |
+
total_size_in_bytes = int(response.headers.get("content-length", 0))
|
| 58 |
+
block_size = 1024 # 1 Kibibyte
|
| 59 |
+
progress_bar = tqdm(
|
| 60 |
+
total=total_size_in_bytes, unit="iB", unit_scale=True
|
| 61 |
+
)
|
| 62 |
+
with open(outf, "wb") as file:
|
| 63 |
+
for data in response.iter_content(block_size):
|
| 64 |
+
progress_bar.update(len(data))
|
| 65 |
+
file.write(data)
|
| 66 |
+
progress_bar.close()
|
| 67 |
+
if total_size_in_bytes != 0 and progress_bar.n != total_size_in_bytes:
|
| 68 |
+
print("ERROR, something went wrong")
|
| 69 |
+
print(f"Downloaded successfully to {outf}")
|
| 70 |
+
p_ckpt = outf
|
| 71 |
+
sd = torch.load(p_ckpt, map_location="cpu")
|
| 72 |
+
unet_lora_config = LoraConfig(
|
| 73 |
+
r=sd["rank_unet"],
|
| 74 |
+
init_lora_weights="gaussian",
|
| 75 |
+
target_modules=sd["unet_lora_target_modules"],
|
| 76 |
+
)
|
| 77 |
+
|
| 78 |
+
if name == "sketch_to_image_stochastic":
|
| 79 |
+
# download from url
|
| 80 |
+
url = "https://www.cs.cmu.edu/~img2img-turbo/models/sketch_to_image_stochastic_lora.pkl"
|
| 81 |
+
os.makedirs(ckpt_folder, exist_ok=True)
|
| 82 |
+
outf = os.path.join(ckpt_folder, "sketch_to_image_stochastic_lora.pkl")
|
| 83 |
+
if not os.path.exists(outf):
|
| 84 |
+
print(f"Downloading checkpoint to {outf}")
|
| 85 |
+
response = requests.get(url, stream=True)
|
| 86 |
+
total_size_in_bytes = int(response.headers.get("content-length", 0))
|
| 87 |
+
block_size = 1024 # 1 Kibibyte
|
| 88 |
+
progress_bar = tqdm(
|
| 89 |
+
total=total_size_in_bytes, unit="iB", unit_scale=True
|
| 90 |
+
)
|
| 91 |
+
with open(outf, "wb") as file:
|
| 92 |
+
for data in response.iter_content(block_size):
|
| 93 |
+
progress_bar.update(len(data))
|
| 94 |
+
file.write(data)
|
| 95 |
+
progress_bar.close()
|
| 96 |
+
if total_size_in_bytes != 0 and progress_bar.n != total_size_in_bytes:
|
| 97 |
+
print("ERROR, something went wrong")
|
| 98 |
+
print(f"Downloaded successfully to {outf}")
|
| 99 |
+
p_ckpt = outf
|
| 100 |
+
sd = torch.load(p_ckpt, map_location="cpu")
|
| 101 |
+
unet_lora_config = LoraConfig(
|
| 102 |
+
r=sd["rank_unet"],
|
| 103 |
+
init_lora_weights="gaussian",
|
| 104 |
+
target_modules=sd["unet_lora_target_modules"],
|
| 105 |
+
)
|
| 106 |
+
convin_pretrained = copy.deepcopy(unet.conv_in)
|
| 107 |
+
unet.conv_in = TwinConv(convin_pretrained, unet.conv_in)
|
| 108 |
+
|
| 109 |
+
vae.encoder.forward = my_vae_encoder_fwd.__get__(
|
| 110 |
+
vae.encoder, vae.encoder.__class__
|
| 111 |
+
)
|
| 112 |
+
vae.decoder.forward = my_vae_decoder_fwd.__get__(
|
| 113 |
+
vae.decoder, vae.decoder.__class__
|
| 114 |
+
)
|
| 115 |
+
# add the skip connection convs
|
| 116 |
+
vae.decoder.skip_conv_1 = torch.nn.Conv2d(
|
| 117 |
+
512, 512, kernel_size=(1, 1), stride=(1, 1), bias=False
|
| 118 |
+
).cuda()
|
| 119 |
+
vae.decoder.skip_conv_2 = torch.nn.Conv2d(
|
| 120 |
+
256, 512, kernel_size=(1, 1), stride=(1, 1), bias=False
|
| 121 |
+
).cuda()
|
| 122 |
+
vae.decoder.skip_conv_3 = torch.nn.Conv2d(
|
| 123 |
+
128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False
|
| 124 |
+
).cuda()
|
| 125 |
+
vae.decoder.skip_conv_4 = torch.nn.Conv2d(
|
| 126 |
+
128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
|
| 127 |
+
).cuda()
|
| 128 |
+
vae_lora_config = LoraConfig(
|
| 129 |
+
r=sd["rank_vae"],
|
| 130 |
+
init_lora_weights="gaussian",
|
| 131 |
+
target_modules=sd["vae_lora_target_modules"],
|
| 132 |
+
)
|
| 133 |
+
vae.decoder.ignore_skip = False
|
| 134 |
+
vae.add_adapter(vae_lora_config, adapter_name="vae_skip")
|
| 135 |
+
unet.add_adapter(unet_lora_config)
|
| 136 |
+
_sd_unet = unet.state_dict()
|
| 137 |
+
for k in sd["state_dict_unet"]:
|
| 138 |
+
_sd_unet[k] = sd["state_dict_unet"][k]
|
| 139 |
+
unet.load_state_dict(_sd_unet)
|
| 140 |
+
unet.enable_xformers_memory_efficient_attention()
|
| 141 |
+
_sd_vae = vae.state_dict()
|
| 142 |
+
for k in sd["state_dict_vae"]:
|
| 143 |
+
_sd_vae[k] = sd["state_dict_vae"][k]
|
| 144 |
+
vae.load_state_dict(_sd_vae)
|
| 145 |
+
unet.to("cuda")
|
| 146 |
+
vae.to("cuda")
|
| 147 |
+
unet.eval()
|
| 148 |
+
vae.eval()
|
| 149 |
+
self.unet, self.vae = unet, vae
|
| 150 |
+
self.vae.decoder.gamma = 1
|
| 151 |
+
self.timesteps = torch.tensor([999], device="cuda").long()
|
| 152 |
+
self.last_prompt = ""
|
| 153 |
+
self.caption_enc = None
|
| 154 |
+
self.device = "cuda"
|
| 155 |
+
|
| 156 |
+
def forward(self, c_t, prompt, deterministic=True, r=1.0, noise_map=1.0):
|
| 157 |
+
# encode the text prompt
|
| 158 |
+
if prompt != self.last_prompt:
|
| 159 |
+
caption_tokens = self.tokenizer(
|
| 160 |
+
prompt,
|
| 161 |
+
max_length=self.tokenizer.model_max_length,
|
| 162 |
+
padding="max_length",
|
| 163 |
+
truncation=True,
|
| 164 |
+
return_tensors="pt",
|
| 165 |
+
).input_ids.cuda()
|
| 166 |
+
caption_enc = self.text_encoder(caption_tokens)[0]
|
| 167 |
+
self.caption_enc = caption_enc
|
| 168 |
+
self.last_prompt = prompt
|
| 169 |
+
|
| 170 |
+
if deterministic:
|
| 171 |
+
encoded_control = (
|
| 172 |
+
self.vae.encode(c_t).latent_dist.sample()
|
| 173 |
+
* self.vae.config.scaling_factor
|
| 174 |
+
)
|
| 175 |
+
model_pred = self.unet(
|
| 176 |
+
encoded_control,
|
| 177 |
+
self.timesteps,
|
| 178 |
+
encoder_hidden_states=self.caption_enc,
|
| 179 |
+
).sample
|
| 180 |
+
x_denoised = self.sched.step(
|
| 181 |
+
model_pred, self.timesteps, encoded_control, return_dict=True
|
| 182 |
+
).prev_sample
|
| 183 |
+
self.vae.decoder.incoming_skip_acts = self.vae.encoder.current_down_blocks
|
| 184 |
+
output_image = (
|
| 185 |
+
self.vae.decode(x_denoised / self.vae.config.scaling_factor).sample
|
| 186 |
+
).clamp(-1, 1)
|
| 187 |
+
else:
|
| 188 |
+
# scale the lora weights based on the r value
|
| 189 |
+
self.unet.set_adapters(["default"], weights=[r])
|
| 190 |
+
set_weights_and_activate_adapters(self.vae, ["vae_skip"], [r])
|
| 191 |
+
encoded_control = (
|
| 192 |
+
self.vae.encode(c_t).latent_dist.sample()
|
| 193 |
+
* self.vae.config.scaling_factor
|
| 194 |
+
)
|
| 195 |
+
# combine the input and noise
|
| 196 |
+
unet_input = encoded_control * r + noise_map * (1 - r)
|
| 197 |
+
self.unet.conv_in.r = r
|
| 198 |
+
unet_output = self.unet(
|
| 199 |
+
unet_input,
|
| 200 |
+
self.timesteps,
|
| 201 |
+
encoder_hidden_states=self.caption_enc,
|
| 202 |
+
).sample
|
| 203 |
+
self.unet.conv_in.r = None
|
| 204 |
+
x_denoised = self.sched.step(
|
| 205 |
+
unet_output, self.timesteps, unet_input, return_dict=True
|
| 206 |
+
).prev_sample
|
| 207 |
+
self.vae.decoder.incoming_skip_acts = self.vae.encoder.current_down_blocks
|
| 208 |
+
self.vae.decoder.gamma = r
|
| 209 |
+
output_image = (
|
| 210 |
+
self.vae.decode(x_denoised / self.vae.config.scaling_factor).sample
|
| 211 |
+
).clamp(-1, 1)
|
| 212 |
+
return output_image
|
server/pipelines/pix2pixTurbo.py
ADDED
|
@@ -0,0 +1,127 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from torchvision import transforms
|
| 3 |
+
|
| 4 |
+
from config import Args
|
| 5 |
+
from pydantic import BaseModel, Field
|
| 6 |
+
from PIL import Image
|
| 7 |
+
from pipelines.pix2pix.pix2pix_turbo import Pix2Pix_Turbo
|
| 8 |
+
from pipelines.utils.canny_gpu import SobelOperator
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
default_prompt = "close-up photo of the joker"
|
| 12 |
+
page_content = """
|
| 13 |
+
<h1 class="text-3xl font-bold">Real-Time pix2pix_turbo</h1>
|
| 14 |
+
<h3 class="text-xl font-bold">pix2pix turbo</h3>
|
| 15 |
+
<p class="text-sm">
|
| 16 |
+
This demo showcases
|
| 17 |
+
<a
|
| 18 |
+
href="https://github.com/GaParmar/img2img-turbo"
|
| 19 |
+
target="_blank"
|
| 20 |
+
class="text-blue-500 underline hover:no-underline">One-Step Image Translation with Text-to-Image Models
|
| 21 |
+
</a>
|
| 22 |
+
</p>
|
| 23 |
+
"""
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
class Pipeline:
|
| 27 |
+
class Info(BaseModel):
|
| 28 |
+
name: str = "img2img"
|
| 29 |
+
title: str = "Image-to-Image SDXL"
|
| 30 |
+
description: str = "Generates an image from a text prompt"
|
| 31 |
+
input_mode: str = "image"
|
| 32 |
+
page_content: str = page_content
|
| 33 |
+
|
| 34 |
+
class InputParams(BaseModel):
|
| 35 |
+
prompt: str = Field(
|
| 36 |
+
default_prompt,
|
| 37 |
+
title="Prompt",
|
| 38 |
+
field="textarea",
|
| 39 |
+
id="prompt",
|
| 40 |
+
)
|
| 41 |
+
|
| 42 |
+
width: int = Field(
|
| 43 |
+
512, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
|
| 44 |
+
)
|
| 45 |
+
height: int = Field(
|
| 46 |
+
512, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
|
| 47 |
+
)
|
| 48 |
+
strength: float = Field(
|
| 49 |
+
1.0,
|
| 50 |
+
min=0.01,
|
| 51 |
+
max=10.0,
|
| 52 |
+
step=0.001,
|
| 53 |
+
title="Strength",
|
| 54 |
+
field="range",
|
| 55 |
+
hide=True,
|
| 56 |
+
id="strength",
|
| 57 |
+
)
|
| 58 |
+
deterministic: bool = Field(
|
| 59 |
+
True,
|
| 60 |
+
hide=True,
|
| 61 |
+
title="Deterministic",
|
| 62 |
+
field="checkbox",
|
| 63 |
+
id="deterministic",
|
| 64 |
+
)
|
| 65 |
+
canny_low_threshold: float = Field(
|
| 66 |
+
0.31,
|
| 67 |
+
min=0,
|
| 68 |
+
max=1.0,
|
| 69 |
+
step=0.001,
|
| 70 |
+
title="Canny Low Threshold",
|
| 71 |
+
field="range",
|
| 72 |
+
hide=True,
|
| 73 |
+
id="canny_low_threshold",
|
| 74 |
+
)
|
| 75 |
+
canny_high_threshold: float = Field(
|
| 76 |
+
0.125,
|
| 77 |
+
min=0,
|
| 78 |
+
max=1.0,
|
| 79 |
+
step=0.001,
|
| 80 |
+
title="Canny High Threshold",
|
| 81 |
+
field="range",
|
| 82 |
+
hide=True,
|
| 83 |
+
id="canny_high_threshold",
|
| 84 |
+
)
|
| 85 |
+
debug_canny: bool = Field(
|
| 86 |
+
False,
|
| 87 |
+
title="Debug Canny",
|
| 88 |
+
field="checkbox",
|
| 89 |
+
hide=True,
|
| 90 |
+
id="debug_canny",
|
| 91 |
+
)
|
| 92 |
+
|
| 93 |
+
def __init__(self, args: Args, device: torch.device, torch_dtype: torch.dtype):
|
| 94 |
+
self.model = Pix2Pix_Turbo("edge_to_image")
|
| 95 |
+
self.canny_torch = SobelOperator(device=device)
|
| 96 |
+
self.device = device
|
| 97 |
+
|
| 98 |
+
def predict(self, params: "Pipeline.InputParams") -> Image.Image:
|
| 99 |
+
# generator = torch.manual_seed(params.seed)
|
| 100 |
+
# pipe = self.pipes[params.base_model_id]
|
| 101 |
+
|
| 102 |
+
canny_pil, canny_tensor = self.canny_torch(
|
| 103 |
+
params.image,
|
| 104 |
+
params.canny_low_threshold,
|
| 105 |
+
params.canny_high_threshold,
|
| 106 |
+
output_type="pil,tensor",
|
| 107 |
+
)
|
| 108 |
+
|
| 109 |
+
with torch.no_grad():
|
| 110 |
+
canny_tensor = torch.cat((canny_tensor, canny_tensor, canny_tensor), dim=1)
|
| 111 |
+
output_image = self.model(
|
| 112 |
+
canny_tensor,
|
| 113 |
+
params.prompt,
|
| 114 |
+
params.deterministic,
|
| 115 |
+
params.strength,
|
| 116 |
+
)
|
| 117 |
+
output_pil = transforms.ToPILImage()(output_image[0].cpu() * 0.5 + 0.5)
|
| 118 |
+
|
| 119 |
+
result_image = output_pil
|
| 120 |
+
if params.debug_canny:
|
| 121 |
+
# paste control_image on top of result_image
|
| 122 |
+
w0, h0 = (200, 200)
|
| 123 |
+
control_image = canny_pil.resize((w0, h0))
|
| 124 |
+
w1, h1 = result_image.size
|
| 125 |
+
result_image.paste(control_image, (w1 - w0, h1 - h0))
|
| 126 |
+
|
| 127 |
+
return result_image
|
server/pipelines/utils/canny_gpu.py
CHANGED
|
@@ -3,6 +3,7 @@ import torch.nn as nn
|
|
| 3 |
from torchvision.transforms import ToTensor, ToPILImage
|
| 4 |
from PIL import Image
|
| 5 |
|
|
|
|
| 6 |
class SobelOperator(nn.Module):
|
| 7 |
def __init__(self, device="cuda"):
|
| 8 |
super(SobelOperator, self).__init__()
|
|
@@ -25,7 +26,13 @@ class SobelOperator(nn.Module):
|
|
| 25 |
self.edge_conv_y.weight = nn.Parameter(sobel_kernel_y.view((1, 1, 3, 3)))
|
| 26 |
|
| 27 |
@torch.no_grad()
|
| 28 |
-
def forward(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
# Convert PIL image to PyTorch tensor
|
| 30 |
image_gray = image.convert("L")
|
| 31 |
image_tensor = ToTensor()(image_gray).unsqueeze(0).to(self.device)
|
|
@@ -41,4 +48,9 @@ class SobelOperator(nn.Module):
|
|
| 41 |
edge[edge <= low_threshold] = 0.0
|
| 42 |
|
| 43 |
# Convert the result back to a PIL image
|
| 44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
from torchvision.transforms import ToTensor, ToPILImage
|
| 4 |
from PIL import Image
|
| 5 |
|
| 6 |
+
|
| 7 |
class SobelOperator(nn.Module):
|
| 8 |
def __init__(self, device="cuda"):
|
| 9 |
super(SobelOperator, self).__init__()
|
|
|
|
| 26 |
self.edge_conv_y.weight = nn.Parameter(sobel_kernel_y.view((1, 1, 3, 3)))
|
| 27 |
|
| 28 |
@torch.no_grad()
|
| 29 |
+
def forward(
|
| 30 |
+
self,
|
| 31 |
+
image: Image.Image,
|
| 32 |
+
low_threshold: float,
|
| 33 |
+
high_threshold: float,
|
| 34 |
+
output_type="pil",
|
| 35 |
+
) -> Image.Image | torch.Tensor | tuple[Image.Image, torch.Tensor]:
|
| 36 |
# Convert PIL image to PyTorch tensor
|
| 37 |
image_gray = image.convert("L")
|
| 38 |
image_tensor = ToTensor()(image_gray).unsqueeze(0).to(self.device)
|
|
|
|
| 48 |
edge[edge <= low_threshold] = 0.0
|
| 49 |
|
| 50 |
# Convert the result back to a PIL image
|
| 51 |
+
if output_type == "pil":
|
| 52 |
+
return ToPILImage()(edge.squeeze(0).cpu())
|
| 53 |
+
elif output_type == "tensor":
|
| 54 |
+
return edge
|
| 55 |
+
elif output_type == "pil,tensor":
|
| 56 |
+
return ToPILImage()(edge.squeeze(0).cpu()), edge
|
server/requirements.txt
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
-
diffusers==0.
|
| 2 |
transformers==4.36.2
|
| 3 |
--extra-index-url https://download.pytorch.org/whl/cu121;
|
| 4 |
torch==2.2.0
|
|
@@ -8,10 +8,11 @@ Pillow==10.2.0
|
|
| 8 |
accelerate==0.25.0
|
| 9 |
compel==2.0.2
|
| 10 |
controlnet-aux==0.0.7
|
| 11 |
-
peft==0.
|
| 12 |
xformers; sys_platform != 'darwin' or platform_machine != 'arm64'
|
| 13 |
markdown2
|
| 14 |
safetensors
|
| 15 |
stable_fast @ https://github.com/chengzeyi/stable-fast/releases/download/v1.0.4/stable_fast-1.0.4+torch220cu121-cp310-cp310-manylinux2014_x86_64.whl ; sys_platform != 'darwin' or platform_machine != 'arm64'
|
| 16 |
oneflow @ https://github.com/siliconflow/oneflow_releases/releases/download/community_cu121/oneflow-0.9.1.dev20240316+cu121-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl ; sys_platform != 'darwin' or platform_machine != 'arm64'
|
| 17 |
-
onediff @ git+https://github.com/siliconflow/onediff.git@main#egg=onediff ; sys_platform != 'darwin' or platform_machine != 'arm64'
|
|
|
|
|
|
| 1 |
+
diffusers==0.25.1
|
| 2 |
transformers==4.36.2
|
| 3 |
--extra-index-url https://download.pytorch.org/whl/cu121;
|
| 4 |
torch==2.2.0
|
|
|
|
| 8 |
accelerate==0.25.0
|
| 9 |
compel==2.0.2
|
| 10 |
controlnet-aux==0.0.7
|
| 11 |
+
peft==0.9.0
|
| 12 |
xformers; sys_platform != 'darwin' or platform_machine != 'arm64'
|
| 13 |
markdown2
|
| 14 |
safetensors
|
| 15 |
stable_fast @ https://github.com/chengzeyi/stable-fast/releases/download/v1.0.4/stable_fast-1.0.4+torch220cu121-cp310-cp310-manylinux2014_x86_64.whl ; sys_platform != 'darwin' or platform_machine != 'arm64'
|
| 16 |
oneflow @ https://github.com/siliconflow/oneflow_releases/releases/download/community_cu121/oneflow-0.9.1.dev20240316+cu121-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl ; sys_platform != 'darwin' or platform_machine != 'arm64'
|
| 17 |
+
onediff @ git+https://github.com/siliconflow/onediff.git@main#egg=onediff ; sys_platform != 'darwin' or platform_machine != 'arm64'
|
| 18 |
+
setuptools
|