Spaces:
Runtime error
Runtime error
init commit
Browse files- Marc Allante.bin +3 -0
- app.py +69 -0
- birb-style.bin +3 -0
- hanfu-anime-style.bin +3 -0
- hitokomoru-style.bin +3 -0
- illustration_style.bin +3 -0
- image_generator.py +381 -0
- line-art.bin +3 -0
- loss.py +64 -0
- midjourney-style.bin +3 -0
- utils.py +30 -0
Marc Allante.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:2b0496315f14f212535f9350c3dbf05787ac50a78465d4be2f39a1ba373e4968
|
3 |
+
size 3819
|
app.py
ADDED
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
|
3 |
+
import os
|
4 |
+
import torch
|
5 |
+
from image_generator import generate_image_per_prompt_style
|
6 |
+
|
7 |
+
torch.manual_seed(11)
|
8 |
+
|
9 |
+
|
10 |
+
# Set device
|
11 |
+
torch_device = (
|
12 |
+
"cuda"
|
13 |
+
if torch.cuda.is_available()
|
14 |
+
else "mps"
|
15 |
+
if torch.backends.mps.is_available()
|
16 |
+
else "cpu"
|
17 |
+
)
|
18 |
+
if "mps" == torch_device:
|
19 |
+
os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
|
20 |
+
# Define Interface
|
21 |
+
title = "Generative Art - Stable Diffusion with Styles and additional guidance"
|
22 |
+
|
23 |
+
gr_interface = gr.Interface(
|
24 |
+
generate_image_per_prompt_style,
|
25 |
+
inputs=[
|
26 |
+
gr.Textbox("cat running", label="Prompt"),
|
27 |
+
gr.Dropdown(
|
28 |
+
[
|
29 |
+
"illustration_style",
|
30 |
+
"line-art",
|
31 |
+
"hitokomoru-style",
|
32 |
+
"midjourney-style",
|
33 |
+
"hanfu-anime-style",
|
34 |
+
"birb-style",
|
35 |
+
"style-of-marc-allante",
|
36 |
+
],
|
37 |
+
value="birb-style",
|
38 |
+
label="Pre-trained Styles",
|
39 |
+
),
|
40 |
+
gr.Dropdown(
|
41 |
+
[
|
42 |
+
"blue_loss",
|
43 |
+
"cosine_loss",
|
44 |
+
],
|
45 |
+
value="cosine_loss",
|
46 |
+
label="Additional guidance for image generation",
|
47 |
+
),
|
48 |
+
gr.Textbox("on a city road", label="Additional Prompt"),
|
49 |
+
],
|
50 |
+
outputs=[
|
51 |
+
gr.Gallery(
|
52 |
+
label="Generated images",
|
53 |
+
show_label=False,
|
54 |
+
elem_id="gallery",
|
55 |
+
columns=[2],
|
56 |
+
rows=[2],
|
57 |
+
object_fit="contain",
|
58 |
+
height="auto",
|
59 |
+
)
|
60 |
+
],
|
61 |
+
title=title,
|
62 |
+
examples=[
|
63 |
+
["A flying bird", "illustration_style", "blue_loss", ""],
|
64 |
+
["cat running", "on a city road", "cosine_loss", ""]
|
65 |
+
]
|
66 |
+
)
|
67 |
+
gr_interface.launch(debug=True)
|
68 |
+
|
69 |
+
|
birb-style.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f2e23a8f2d3628ed77acb8151751ecd4efc4017e8da86bc29af10f855ca308d9
|
3 |
+
size 3819
|
hanfu-anime-style.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:18ee85c31cff7a0ab35f90af24fbf1a4ab8a9960ab041511e386d5990953e050
|
3 |
+
size 3819
|
hitokomoru-style.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f81a9c575e329e08a24e08f47ae73c5b50dec4bcb557974552549b45e2d1b0d4
|
3 |
+
size 3819
|
illustration_style.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:44d65046c071e37f75f31a7a81a34c50a96080e8a3aedc7cda1094dae5d385f0
|
3 |
+
size 3819
|
image_generator.py
ADDED
@@ -0,0 +1,381 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from pathlib import Path
|
3 |
+
import torch
|
4 |
+
from diffusers import AutoencoderKL, LMSDiscreteScheduler, UNet2DConditionModel
|
5 |
+
from tqdm.auto import tqdm
|
6 |
+
from transformers import CLIPTextModel, CLIPTokenizer, logging
|
7 |
+
|
8 |
+
from utils import load_embedding_bin, set_timesteps, latents_to_pil
|
9 |
+
from loss import blue_loss, cosine_loss
|
10 |
+
from matplotlib import pyplot as plt
|
11 |
+
from pathlib import Path
|
12 |
+
|
13 |
+
torch.manual_seed(11)
|
14 |
+
logging.set_verbosity_error()
|
15 |
+
|
16 |
+
# Set device
|
17 |
+
torch_device = (
|
18 |
+
"cuda"
|
19 |
+
if torch.cuda.is_available()
|
20 |
+
else "mps"
|
21 |
+
if torch.backends.mps.is_available()
|
22 |
+
else "cpu"
|
23 |
+
)
|
24 |
+
if "mps" == torch_device:
|
25 |
+
os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
|
26 |
+
# Style embeddings
|
27 |
+
STYLE_EMBEDDINGS = {
|
28 |
+
"illustration-style": "illustration_style.bin",
|
29 |
+
"line-art": "line-art.bin",
|
30 |
+
"hitokomoru-style": "hitokomoru-style.bin",
|
31 |
+
"midjourney-style": "midjourney-style.bin",
|
32 |
+
"hanfu-anime-style": "hanfu-anime-style.bin",
|
33 |
+
"birb-style": "birb-style.bin",
|
34 |
+
"style-of-marc-allante": "Marc Allante.bin",
|
35 |
+
}
|
36 |
+
LOSS = {"blue_loss": blue_loss,
|
37 |
+
"cosine_loss": cosine_loss}
|
38 |
+
STYLE_SEEDS = [11, 56, 110, 65, 5, 29, 47]
|
39 |
+
# Load the autoencoder model which will be used to decode the latents into image space.
|
40 |
+
vae = AutoencoderKL.from_pretrained(
|
41 |
+
"CompVis/stable-diffusion-v1-4", subfolder="vae"
|
42 |
+
).to(torch_device)
|
43 |
+
#
|
44 |
+
# # Load the tokenizer and text encoder to tokenize and encode the text.
|
45 |
+
tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
|
46 |
+
text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14").to(
|
47 |
+
torch_device
|
48 |
+
)
|
49 |
+
#
|
50 |
+
# # The UNet model for generating the latents.
|
51 |
+
unet = UNet2DConditionModel.from_pretrained(
|
52 |
+
"CompVis/stable-diffusion-v1-4", subfolder="unet"
|
53 |
+
).to(torch_device)
|
54 |
+
#
|
55 |
+
# # The noise scheduler
|
56 |
+
scheduler = LMSDiscreteScheduler(
|
57 |
+
beta_start=0.00085,
|
58 |
+
beta_end=0.012,
|
59 |
+
beta_schedule="scaled_linear",
|
60 |
+
num_train_timesteps=1000,
|
61 |
+
)
|
62 |
+
|
63 |
+
# vae = vae
|
64 |
+
# text_encoder = text_encoder.to(torch_device)
|
65 |
+
unet = unet
|
66 |
+
token_emb_layer = text_encoder.text_model.embeddings.token_embedding
|
67 |
+
pos_emb_layer = text_encoder.text_model.embeddings.position_embedding
|
68 |
+
position_ids = text_encoder.text_model.embeddings.position_ids[:, :77]
|
69 |
+
position_embeddings = pos_emb_layer(position_ids)
|
70 |
+
|
71 |
+
|
72 |
+
def build_causal_attention_mask(bsz, seq_len, dtype):
|
73 |
+
# lazily create causal attention mask, with full attention between the vision tokens
|
74 |
+
# pytorch uses additive attention mask; fill with -inf
|
75 |
+
mask = torch.empty(bsz, seq_len, seq_len, dtype=dtype)
|
76 |
+
mask.fill_(torch.tensor(torch.finfo(dtype).min))
|
77 |
+
mask.triu_(1) # zero out the lower diagonal
|
78 |
+
mask = mask.unsqueeze(1) # expand mask
|
79 |
+
return mask
|
80 |
+
|
81 |
+
|
82 |
+
def get_output_embeds(input_embeddings):
|
83 |
+
# CLIP's text model uses causal mask, so we prepare it here:
|
84 |
+
bsz, seq_len = input_embeddings.shape[:2]
|
85 |
+
causal_attention_mask = build_causal_attention_mask(
|
86 |
+
bsz, seq_len, dtype=input_embeddings.dtype
|
87 |
+
)
|
88 |
+
|
89 |
+
# Getting the output embeddings involves calling the model with passing output_hidden_states=True
|
90 |
+
# so that it doesn't just return the pooled final predictions:
|
91 |
+
encoder_outputs = text_encoder.text_model.encoder(
|
92 |
+
inputs_embeds=input_embeddings,
|
93 |
+
attention_mask=None, # We aren't using an attention mask so that can be None
|
94 |
+
causal_attention_mask=causal_attention_mask.to(torch_device),
|
95 |
+
output_attentions=None,
|
96 |
+
output_hidden_states=True, # We want the output embs not the final output
|
97 |
+
return_dict=None,
|
98 |
+
)
|
99 |
+
|
100 |
+
# We're interested in the output hidden state only
|
101 |
+
output = encoder_outputs[0]
|
102 |
+
|
103 |
+
# There is a final layer norm we need to pass these through
|
104 |
+
output = text_encoder.text_model.final_layer_norm(output)
|
105 |
+
|
106 |
+
# And now they're ready!
|
107 |
+
return output
|
108 |
+
|
109 |
+
|
110 |
+
# Generating an image with these modified embeddings
|
111 |
+
def generate_with_embs(text_embeddings, seed, max_length):
|
112 |
+
height = 512 # default height of Stable Diffusion
|
113 |
+
width = 512 # default width of Stable Diffusion
|
114 |
+
num_inference_steps = 30 # Number of denoising steps
|
115 |
+
guidance_scale = 7.5 # Scale for classifier-free guidance
|
116 |
+
generator = torch.manual_seed(seed)
|
117 |
+
batch_size = 1
|
118 |
+
|
119 |
+
# tokenizer
|
120 |
+
uncond_input = tokenizer(
|
121 |
+
[""] * batch_size,
|
122 |
+
padding="max_length",
|
123 |
+
max_length=max_length,
|
124 |
+
return_tensors="pt",
|
125 |
+
)
|
126 |
+
|
127 |
+
with torch.no_grad():
|
128 |
+
uncond_embeddings = text_encoder(uncond_input.input_ids.to(torch_device))[0]
|
129 |
+
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
130 |
+
|
131 |
+
# Prep Scheduler
|
132 |
+
set_timesteps(scheduler, num_inference_steps)
|
133 |
+
|
134 |
+
# Prep latents
|
135 |
+
# step = " prep_latents "
|
136 |
+
latents = torch.randn(
|
137 |
+
(batch_size, unet.in_channels, height // 8, width // 8),
|
138 |
+
generator=generator,
|
139 |
+
)
|
140 |
+
latents = latents.to(torch_device)
|
141 |
+
latents = latents * scheduler.init_noise_sigma
|
142 |
+
|
143 |
+
# Loop
|
144 |
+
for i, t in tqdm(enumerate(scheduler.timesteps),
|
145 |
+
total=len(scheduler.timesteps)):
|
146 |
+
# expand the latents if we are doing classifier-free guidance to avoid doing two forward passes.
|
147 |
+
latent_model_input = torch.cat([latents] * 2)
|
148 |
+
latent_model_input = scheduler.scale_model_input(latent_model_input, t)
|
149 |
+
|
150 |
+
# predict the noise residual
|
151 |
+
with torch.no_grad():
|
152 |
+
noise_pred = unet(
|
153 |
+
latent_model_input, t, encoder_hidden_states=text_embeddings
|
154 |
+
)["sample"]
|
155 |
+
|
156 |
+
# perform guidance
|
157 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
158 |
+
noise_pred = noise_pred_uncond + guidance_scale * (
|
159 |
+
noise_pred_text - noise_pred_uncond
|
160 |
+
)
|
161 |
+
|
162 |
+
# compute the previous noisy sample x_t -> x_t-1
|
163 |
+
latents = scheduler.step(noise_pred, t, latents).prev_sample
|
164 |
+
|
165 |
+
return latents_to_pil(latents)[0]
|
166 |
+
|
167 |
+
|
168 |
+
def generate_image_from_embeddings(
|
169 |
+
mod_output_embeddings, seed, max_length,
|
170 |
+
loss_selection, additional_prompt):
|
171 |
+
height = 512
|
172 |
+
width = 512
|
173 |
+
num_inference_steps = 50
|
174 |
+
guidance_scale = 8
|
175 |
+
generator = torch.manual_seed(seed)
|
176 |
+
batch_size = 1
|
177 |
+
if loss_selection == "blue_loss":
|
178 |
+
loss_fn = LOSS["blue_loss"]
|
179 |
+
loss_scale = 120
|
180 |
+
else:
|
181 |
+
loss_fn = LOSS["cosine_loss"](additional_prompt)
|
182 |
+
loss_scale = 20
|
183 |
+
|
184 |
+
# Use the modified_output_embeddings directly
|
185 |
+
text_embeddings = mod_output_embeddings
|
186 |
+
|
187 |
+
uncond_input = tokenizer(
|
188 |
+
[""] * batch_size,
|
189 |
+
padding="max_length",
|
190 |
+
max_length=max_length,
|
191 |
+
return_tensors="pt",
|
192 |
+
)
|
193 |
+
with torch.no_grad():
|
194 |
+
uncond_embeddings = text_encoder(
|
195 |
+
uncond_input.input_ids.to(torch_device))[0]
|
196 |
+
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
197 |
+
|
198 |
+
# Prep Scheduler
|
199 |
+
set_timesteps(scheduler, num_inference_steps)
|
200 |
+
|
201 |
+
# Prep latents
|
202 |
+
latents = torch.randn(
|
203 |
+
(batch_size, unet.config.in_channels, height // 8, width // 8),
|
204 |
+
generator=generator,
|
205 |
+
)
|
206 |
+
latents = latents.to(torch_device)
|
207 |
+
latents = latents * scheduler.init_noise_sigma
|
208 |
+
|
209 |
+
# Loop
|
210 |
+
for i, t in tqdm(enumerate(scheduler.timesteps),
|
211 |
+
total=len(scheduler.timesteps)):
|
212 |
+
# expand the latents if we are doing classifier-free guidance to avoid doing two forward passes.
|
213 |
+
latent_model_input = torch.cat([latents] * 2)
|
214 |
+
sigma = scheduler.sigmas[i]
|
215 |
+
latent_model_input = scheduler.scale_model_input(latent_model_input, t)
|
216 |
+
|
217 |
+
# predict the noise residual
|
218 |
+
with torch.no_grad():
|
219 |
+
noise_pred = unet(
|
220 |
+
latent_model_input, t, encoder_hidden_states=text_embeddings
|
221 |
+
)["sample"]
|
222 |
+
|
223 |
+
# perform CFG
|
224 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
225 |
+
noise_pred = noise_pred_uncond + guidance_scale * (
|
226 |
+
noise_pred_text - noise_pred_uncond
|
227 |
+
)
|
228 |
+
|
229 |
+
#### ADDITIONAL GUIDANCE ###
|
230 |
+
if i % 2 == 0:
|
231 |
+
# Requires grad on the latents
|
232 |
+
latents = latents.detach().requires_grad_()
|
233 |
+
|
234 |
+
# Get the predicted x0:
|
235 |
+
# latents_x0 = latents - sigma * noise_pred
|
236 |
+
latents_x0 = scheduler.step(noise_pred, t, latents).pred_original_sample
|
237 |
+
scheduler._step_index -= 1
|
238 |
+
# Decode to image space
|
239 |
+
denoised_images = (
|
240 |
+
vae.decode((1 / 0.18215) * latents_x0).sample / 2 + 0.5
|
241 |
+
) # range (0, 1)
|
242 |
+
|
243 |
+
# Calculate loss
|
244 |
+
loss = loss_fn(denoised_images) * loss_scale
|
245 |
+
|
246 |
+
# Occasionally print it out
|
247 |
+
if i % 10 == 0:
|
248 |
+
print(i, "loss:", loss.item())
|
249 |
+
|
250 |
+
# Get gradient
|
251 |
+
cond_grad = torch.autograd.grad(loss, latents)[0]
|
252 |
+
|
253 |
+
# Modify the latents based on this gradient
|
254 |
+
latents = latents.detach() - cond_grad * sigma**2
|
255 |
+
|
256 |
+
# Now step with scheduler
|
257 |
+
latents = scheduler.step(noise_pred, t, latents).prev_sample
|
258 |
+
|
259 |
+
return latents_to_pil(latents)[0]
|
260 |
+
|
261 |
+
|
262 |
+
def generate_image_per_style(prompt, style_embed, style_seed, style_embedding_key):
|
263 |
+
modified_output_embeddings = None
|
264 |
+
gen_out_style_image = None
|
265 |
+
max_length = 0
|
266 |
+
|
267 |
+
# Tokenize
|
268 |
+
text_input = tokenizer(
|
269 |
+
prompt,
|
270 |
+
padding="max_length",
|
271 |
+
max_length=tokenizer.model_max_length,
|
272 |
+
truncation=True,
|
273 |
+
return_tensors="pt",
|
274 |
+
)
|
275 |
+
input_ids = text_input.input_ids.to(torch_device)
|
276 |
+
|
277 |
+
# Get token embeddings
|
278 |
+
token_embeddings = token_emb_layer(input_ids)
|
279 |
+
|
280 |
+
replacement_token_embedding = style_embed[style_embedding_key]
|
281 |
+
|
282 |
+
# Insert this into the token embeddings
|
283 |
+
token_embeddings[
|
284 |
+
0, torch.where(input_ids[0] == 6829)[0]
|
285 |
+
] = replacement_token_embedding.to(torch_device)
|
286 |
+
|
287 |
+
# Combine with pos embs
|
288 |
+
input_embeddings = token_embeddings + position_embeddings
|
289 |
+
|
290 |
+
# Feed through to get final output embs
|
291 |
+
modified_output_embeddings = get_output_embeds(input_embeddings)
|
292 |
+
|
293 |
+
# And generate an image with this:
|
294 |
+
max_length = text_input.input_ids.shape[-1]
|
295 |
+
|
296 |
+
gen_out_style_image = generate_with_embs(
|
297 |
+
modified_output_embeddings, style_seed, max_length
|
298 |
+
)
|
299 |
+
|
300 |
+
return gen_out_style_image
|
301 |
+
|
302 |
+
|
303 |
+
def generate_image_per_loss(
|
304 |
+
prompt, style_embed, style_seed, style_embedding_key,
|
305 |
+
loss, additional_prompt
|
306 |
+
):
|
307 |
+
gen_out_loss_image = None
|
308 |
+
|
309 |
+
# Tokenize
|
310 |
+
text_input = tokenizer(
|
311 |
+
prompt,
|
312 |
+
padding="max_length",
|
313 |
+
max_length=tokenizer.model_max_length,
|
314 |
+
truncation=True,
|
315 |
+
return_tensors="pt",
|
316 |
+
)
|
317 |
+
input_ids = text_input.input_ids.to(torch_device)
|
318 |
+
|
319 |
+
# Get token embeddings
|
320 |
+
token_embeddings = token_emb_layer(input_ids)
|
321 |
+
|
322 |
+
replacement_token_embedding = style_embed[style_embedding_key].to(torch_device)
|
323 |
+
|
324 |
+
# Insert this into the token embeddings
|
325 |
+
token_embeddings[
|
326 |
+
0, torch.where(input_ids[0] == 6829)[0]
|
327 |
+
] = replacement_token_embedding
|
328 |
+
|
329 |
+
# Combine with pos embs
|
330 |
+
input_embeddings = token_embeddings + position_embeddings
|
331 |
+
modified_output_embeddings = get_output_embeds(input_embeddings)
|
332 |
+
|
333 |
+
# max_length = tokenizer.model_max_length
|
334 |
+
|
335 |
+
max_length = text_input.input_ids.shape[-1]
|
336 |
+
gen_out_loss_image = generate_image_from_embeddings(
|
337 |
+
modified_output_embeddings, style_seed, max_length,
|
338 |
+
loss, additional_prompt
|
339 |
+
)
|
340 |
+
|
341 |
+
return gen_out_loss_image
|
342 |
+
|
343 |
+
|
344 |
+
def generate_image_per_prompt_style(text_in, style_in,
|
345 |
+
loss, additional_prompt):
|
346 |
+
gen_style_image = None
|
347 |
+
gen_loss_image = None
|
348 |
+
STYLE_KEYS = []
|
349 |
+
style_key = ""
|
350 |
+
|
351 |
+
if style_in not in STYLE_EMBEDDINGS:
|
352 |
+
raise ValueError(
|
353 |
+
f"Unknown style: {style_in}. Available styles are: {', '.join(STYLE_EMBEDDINGS.keys())}"
|
354 |
+
)
|
355 |
+
|
356 |
+
STYLE_SEEDS = [32, 64, 128, 16, 8, 96]
|
357 |
+
STYLE_KEYS = list(STYLE_EMBEDDINGS.keys())
|
358 |
+
print(f"prompt: {text_in}")
|
359 |
+
print(f"style: {style_in}")
|
360 |
+
|
361 |
+
idx = STYLE_KEYS.index(style_in)
|
362 |
+
style_file = STYLE_EMBEDDINGS[style_in]
|
363 |
+
print(f"style_file: {style_file}")
|
364 |
+
|
365 |
+
prompt = text_in
|
366 |
+
|
367 |
+
style_seed = STYLE_SEEDS[idx]
|
368 |
+
|
369 |
+
style_key = Path(style_file).stem
|
370 |
+
style_key = style_key.replace("_", "-")
|
371 |
+
print(style_key, STYLE_KEYS, style_file)
|
372 |
+
|
373 |
+
file_path = os.path.join(os.getcwd(), style_file)
|
374 |
+
embedding = load_embedding_bin(file_path)
|
375 |
+
style_key = f"<{style_key}>"
|
376 |
+
|
377 |
+
gen_style_image = generate_image_per_style(prompt, embedding, style_seed, style_key)
|
378 |
+
|
379 |
+
gen_loss_image = generate_image_per_loss(prompt, embedding, style_seed, style_key, loss, additional_prompt)
|
380 |
+
|
381 |
+
return [gen_style_image, gen_loss_image]
|
line-art.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:0528436ec2228c659e0cf1316e713345bc97a3d88294f1a2987a3505d220e770
|
3 |
+
size 3819
|
loss.py
ADDED
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from torchvision.transforms import v2
|
5 |
+
from transformers import CLIPTextModel, CLIPTokenizer, \
|
6 |
+
CLIPProcessor, CLIPVisionModelWithProjection, CLIPTextModelWithProjection
|
7 |
+
|
8 |
+
import os
|
9 |
+
# from image_generator import get_output_embeds, position_embeddings
|
10 |
+
|
11 |
+
|
12 |
+
# Set device
|
13 |
+
torch_device = "cuda" if torch.cuda.is_available() else "mps" \
|
14 |
+
if torch.backends.mps.is_available() else "cpu"
|
15 |
+
|
16 |
+
if "mps" == torch_device:
|
17 |
+
os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = "1"
|
18 |
+
|
19 |
+
# Load the tokenizer and text encoder to tokenize and encode the text.
|
20 |
+
clip_model_name = "openai/clip-vit-large-patch14"
|
21 |
+
tokenizer = CLIPTokenizer.from_pretrained(clip_model_name)
|
22 |
+
text_encoder = CLIPTextModel.from_pretrained(clip_model_name).to(torch_device);
|
23 |
+
vision_encoder = CLIPVisionModelWithProjection.from_pretrained(clip_model_name).to(torch_device);
|
24 |
+
processor = CLIPProcessor.from_pretrained(clip_model_name)
|
25 |
+
|
26 |
+
# # additional textual prompt
|
27 |
+
def get_text_embed(prompt = "on a mountain"):
|
28 |
+
inputs = processor(text=prompt,
|
29 |
+
return_tensors="pt",
|
30 |
+
padding=True)
|
31 |
+
with torch.no_grad():
|
32 |
+
text_embed = CLIPTextModelWithProjection.from_pretrained(
|
33 |
+
clip_model_name)(**inputs).text_embeds.to(torch_device)
|
34 |
+
return text_embed
|
35 |
+
|
36 |
+
# def get_text_embed(prompt = "on a mountain"):
|
37 |
+
# text_input = tokenizer([prompt],
|
38 |
+
# padding="max_length",
|
39 |
+
# max_length=tokenizer.model_max_length,
|
40 |
+
# truncation=True,
|
41 |
+
# return_tensors="pt")
|
42 |
+
# with torch.no_grad():
|
43 |
+
# text_embeddings = text_encoder(text_input.input_ids.to(torch_device))[0]
|
44 |
+
# input_embeddings = text_embeddings + position_embeddings.to(torch_device)
|
45 |
+
# modified_output_embeddings = get_output_embeds(input_embeddings)
|
46 |
+
# return modified_output_embeddings
|
47 |
+
|
48 |
+
class cosine_loss(nn.Module):
|
49 |
+
def __init__(self, prompt) -> None:
|
50 |
+
self.text_embed = get_text_embed(prompt)
|
51 |
+
super().__init__()
|
52 |
+
|
53 |
+
def forward(self, gen_image):
|
54 |
+
gen_image_clamped = gen_image.clamp(0, 1).mul(255)
|
55 |
+
resized_image = v2.Resize(224)(gen_image_clamped)
|
56 |
+
image_embed = vision_encoder(resized_image).image_embeds
|
57 |
+
similarity = F.cosine_similarity(self.text_embed, image_embed, dim=1)
|
58 |
+
loss = 1 - similarity.mean()
|
59 |
+
return loss
|
60 |
+
|
61 |
+
def blue_loss(images):
|
62 |
+
# How far are the blue channel values to 0.9:
|
63 |
+
error = torch.abs(images[:,2] - 0.9).mean() # [:,2] -> all images in batch, only the blue channel
|
64 |
+
return error
|
midjourney-style.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4865a5d2ecd012985940748023fd80e4fd299837f1dccedb85ee83be5bb1f957
|
3 |
+
size 3819
|
utils.py
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from PIL import Image
|
3 |
+
from diffusers import AutoencoderKL
|
4 |
+
|
5 |
+
vae = AutoencoderKL.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="vae").to("mps:0")
|
6 |
+
|
7 |
+
def pil_to_latent(input_im):
|
8 |
+
# Single image -> single latent in a batch (so size 1, 4, 64, 64)
|
9 |
+
with torch.no_grad():
|
10 |
+
latent = vae.encode(tfms.ToTensor()(input_im).unsqueeze(0).to(torch_device)*2-1) # Note scaling
|
11 |
+
return 0.18215 * latent.latent_dist.sample()
|
12 |
+
|
13 |
+
def latents_to_pil(latents, torch_device="mps:0"):
|
14 |
+
# bath of latents -> list of images
|
15 |
+
latents = (1 / 0.18215) * latents
|
16 |
+
with torch.no_grad():
|
17 |
+
image = vae.decode(latents.to(torch_device)).sample
|
18 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
19 |
+
image = image.detach().cpu().permute(0, 2, 3, 1).numpy()
|
20 |
+
images = (image * 255).round().astype("uint8")
|
21 |
+
pil_images = [Image.fromarray(image) for image in images]
|
22 |
+
return pil_images
|
23 |
+
|
24 |
+
def load_embedding_bin(path):
|
25 |
+
return torch.load(path)
|
26 |
+
|
27 |
+
# Prep Scheduler
|
28 |
+
def set_timesteps(scheduler, num_inference_steps):
|
29 |
+
scheduler.set_timesteps(num_inference_steps)
|
30 |
+
scheduler.timesteps = scheduler.timesteps.to(torch.float32) # minor fix to ensure MPS compatibility, fixed in diffusers PR 3925
|