Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,260 +1,131 @@
|
|
1 |
import gradio as gr
|
2 |
-
import numpy as np
|
3 |
-
import random
|
4 |
-
import os
|
5 |
-
|
6 |
-
# import spaces #[uncomment to use ZeroGPU]
|
7 |
-
from diffusers import DiffusionPipeline, StableDiffusionPipeline
|
8 |
-
from peft import PeftModel, LoraConfig
|
9 |
import torch
|
10 |
-
from
|
11 |
-
|
|
|
12 |
|
13 |
def get_lora_sd_pipeline(
|
14 |
ckpt_dir='./lora_logos',
|
15 |
base_model_name_or_path=None,
|
16 |
dtype=torch.float16,
|
17 |
adapter_name="default"
|
18 |
-
):
|
19 |
unet_sub_dir = os.path.join(ckpt_dir, "unet")
|
20 |
text_encoder_sub_dir = os.path.join(ckpt_dir, "text_encoder")
|
|
|
21 |
if os.path.exists(text_encoder_sub_dir) and base_model_name_or_path is None:
|
22 |
config = LoraConfig.from_pretrained(text_encoder_sub_dir)
|
23 |
base_model_name_or_path = config.base_model_name_or_path
|
24 |
-
|
25 |
if base_model_name_or_path is None:
|
26 |
raise ValueError("Please specify the base model name or path")
|
27 |
-
|
28 |
pipe = StableDiffusionPipeline.from_pretrained(base_model_name_or_path, torch_dtype=dtype)
|
29 |
pipe.unet = PeftModel.from_pretrained(pipe.unet, unet_sub_dir, adapter_name=adapter_name)
|
30 |
-
|
31 |
if os.path.exists(text_encoder_sub_dir):
|
32 |
-
pipe.text_encoder = PeftModel.from_pretrained(
|
33 |
-
|
34 |
-
)
|
35 |
-
|
36 |
if dtype in (torch.float16, torch.bfloat16):
|
37 |
pipe.unet.half()
|
38 |
pipe.text_encoder.half()
|
|
|
39 |
return pipe
|
40 |
|
41 |
-
def
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
def get_prompt_embeds(prompt_chunks, text_encoder):
|
50 |
-
prompt_embeds = []
|
51 |
-
for chunk in prompt_chunks:
|
52 |
-
chunk_tensor = torch.tensor([chunk]).to(text_encoder.device)
|
53 |
-
with torch.no_grad():
|
54 |
-
embeds = text_encoder(chunk_tensor)[0]
|
55 |
-
prompt_embeds.append(embeds)
|
56 |
-
return torch.cat(prompt_embeds, dim=1)
|
57 |
|
58 |
-
def
|
59 |
max_length = max(prompt_embeds.shape[1], negative_prompt_embeds.shape[1])
|
|
|
|
|
60 |
|
61 |
-
|
62 |
-
padding = target_length - tensor.shape[1]
|
63 |
-
if padding > 0:
|
64 |
-
pad_tensor = torch.zeros(
|
65 |
-
tensor.shape[0], padding, tensor.shape[2], device=tensor.device
|
66 |
-
)
|
67 |
-
tensor = torch.cat([tensor, pad_tensor], dim=1)
|
68 |
-
return tensor
|
69 |
-
|
70 |
-
prompt_embeds = pad_to_max_length(prompt_embeds, max_length)
|
71 |
-
negative_prompt_embeds = pad_to_max_length(negative_prompt_embeds, max_length)
|
72 |
-
|
73 |
-
assert prompt_embeds.shape == negative_prompt_embeds.shape, "Shapes do not match!"
|
74 |
-
return prompt_embeds, negative_prompt_embeds
|
75 |
-
|
76 |
-
def prompts_embeddings(prompt, negative_promt, tokenizer, text_encoder):
|
77 |
-
prompt_chunks = split_prompt(prompt, tokenizer)
|
78 |
-
negative_prompt_chunks = split_prompt(negative_prompt, tokenizer)
|
79 |
-
|
80 |
-
prompt_embeds = get_prompt_embeds(prompt_chunks, text_encoder)
|
81 |
-
negative_prompt_embeds = get_prompt_embeds(negative_prompt_chunks, text_encoder)
|
82 |
-
|
83 |
-
prompt_embeds, negative_prompt_embeds = shape_alignment(prompt_embeds, negative_prompt_embeds)
|
84 |
-
|
85 |
-
return prompt_embeds, negative_prompt_embeds
|
86 |
-
|
87 |
-
|
88 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
89 |
model_id_default = "CompVis/stable-diffusion-v1-4"
|
|
|
90 |
|
91 |
-
|
92 |
-
torch_dtype = torch.float16
|
93 |
-
else:
|
94 |
-
torch_dtype = torch.float32
|
95 |
-
|
96 |
-
|
97 |
-
pipe_default = get_lora_sd_pipeline(
|
98 |
-
ckpt_dir='./lora_logos',
|
99 |
-
base_model_name_or_path=model_id_default,
|
100 |
-
dtype=torch_dtype,
|
101 |
-
)
|
102 |
-
# pipe_default = DiffusionPipeline.from_pretrained(model_id_default, torch_dtype=torch_dtype)
|
103 |
-
pipe_default = pipe_default.to(device)
|
104 |
-
|
105 |
-
MAX_SEED = np.iinfo(np.int32).max
|
106 |
-
MAX_IMAGE_SIZE = 1024
|
107 |
|
108 |
-
|
109 |
-
# @spaces.GPU #[uncomment to use ZeroGPU]
|
110 |
def infer(
|
111 |
-
prompt
|
112 |
-
negative_prompt
|
113 |
-
width
|
114 |
-
height
|
115 |
-
num_inference_steps
|
116 |
-
model_id
|
117 |
-
seed
|
118 |
-
guidance_scale
|
119 |
-
lora_scale
|
120 |
-
|
121 |
-
)
|
122 |
-
|
|
|
|
|
123 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
params = {
|
125 |
-
|
126 |
-
|
127 |
'guidance_scale': guidance_scale,
|
128 |
'num_inference_steps': num_inference_steps,
|
129 |
'width': width,
|
130 |
'height': height,
|
131 |
'generator': generator,
|
132 |
}
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
params['negative_prompt_embeds']=negative_prompt_embeds
|
147 |
-
pipe_default.fuse_lora(lora_scale=lora_scale)
|
148 |
-
image = pipe_default(**params).images[0]
|
149 |
-
|
150 |
-
return image
|
151 |
-
|
152 |
-
css = """
|
153 |
-
#col-container {
|
154 |
-
margin: 0 auto;
|
155 |
-
max-width: 640px;
|
156 |
-
}
|
157 |
-
"""
|
158 |
-
|
159 |
-
with gr.Blocks(css=css) as demo:
|
160 |
-
with gr.Column(elem_id="col-container"):
|
161 |
|
162 |
-
gr.Markdown(" # DEMO Text-to-Image")
|
163 |
-
|
164 |
-
with gr.Row():
|
165 |
-
model_id = gr.Textbox(
|
166 |
-
label="Model ID",
|
167 |
-
max_lines=1,
|
168 |
-
placeholder="Enter model id like 'CompVis/stable-diffusion-v1-4'",
|
169 |
-
value="CompVis/stable-diffusion-v1-4"
|
170 |
-
)
|
171 |
-
|
172 |
-
prompt = gr.Textbox(
|
173 |
-
label="Prompt",
|
174 |
-
max_lines=1,
|
175 |
-
placeholder="Enter your prompt",
|
176 |
-
)
|
177 |
-
|
178 |
-
negative_prompt = gr.Textbox(
|
179 |
-
label="Negative prompt",
|
180 |
-
max_lines=1,
|
181 |
-
placeholder="Enter a negative prompt",
|
182 |
-
)
|
183 |
-
|
184 |
-
with gr.Row():
|
185 |
-
seed = gr.Number(
|
186 |
-
label="Seed",
|
187 |
-
minimum=0,
|
188 |
-
maximum=MAX_SEED,
|
189 |
-
step=1,
|
190 |
-
value=42,
|
191 |
-
)
|
192 |
-
|
193 |
-
with gr.Row():
|
194 |
-
guidance_scale = gr.Slider(
|
195 |
-
label="Guidance scale",
|
196 |
-
minimum=0.0,
|
197 |
-
maximum=10.0,
|
198 |
-
step=0.1,
|
199 |
-
value=7.0,
|
200 |
-
)
|
201 |
-
|
202 |
-
with gr.Row():
|
203 |
-
lora_scale = gr.Slider(
|
204 |
-
label="LoRA scale",
|
205 |
-
minimum=0.0,
|
206 |
-
maximum=1.0,
|
207 |
-
step=0.1,
|
208 |
-
value=0.5,
|
209 |
-
)
|
210 |
-
|
211 |
-
with gr.Row():
|
212 |
-
num_inference_steps = gr.Slider(
|
213 |
-
label="Number of inference steps",
|
214 |
-
minimum=1,
|
215 |
-
maximum=50,
|
216 |
-
step=1,
|
217 |
-
value=20,
|
218 |
-
)
|
219 |
-
|
220 |
with gr.Accordion("Optional Settings", open=False):
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
step=32,
|
227 |
-
value=512,
|
228 |
-
)
|
229 |
-
|
230 |
-
with gr.Row():
|
231 |
-
height = gr.Slider(
|
232 |
-
label="Height",
|
233 |
-
minimum=256,
|
234 |
-
maximum=MAX_IMAGE_SIZE,
|
235 |
-
step=32,
|
236 |
-
value=512,
|
237 |
-
)
|
238 |
-
|
239 |
-
run_button = gr.Button("Run", scale=1, variant="primary")
|
240 |
-
result = gr.Image(label="Result", show_label=False)
|
241 |
|
242 |
-
|
243 |
-
|
244 |
-
fn=infer,
|
245 |
inputs=[
|
246 |
-
prompt,
|
247 |
-
negative_prompt,
|
248 |
-
width,
|
249 |
-
height,
|
250 |
-
num_inference_steps,
|
251 |
-
model_id,
|
252 |
-
|
253 |
-
|
254 |
-
|
255 |
-
|
256 |
-
outputs=[result],
|
257 |
-
)
|
258 |
|
259 |
if __name__ == "__main__":
|
260 |
-
demo.launch()
|
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
import torch
|
3 |
+
from diffusers import StableDiffusionPipeline
|
4 |
+
from peft import PeftModel, LoraConfig
|
5 |
+
import os
|
6 |
|
7 |
def get_lora_sd_pipeline(
|
8 |
ckpt_dir='./lora_logos',
|
9 |
base_model_name_or_path=None,
|
10 |
dtype=torch.float16,
|
11 |
adapter_name="default"
|
12 |
+
):
|
13 |
unet_sub_dir = os.path.join(ckpt_dir, "unet")
|
14 |
text_encoder_sub_dir = os.path.join(ckpt_dir, "text_encoder")
|
15 |
+
|
16 |
if os.path.exists(text_encoder_sub_dir) and base_model_name_or_path is None:
|
17 |
config = LoraConfig.from_pretrained(text_encoder_sub_dir)
|
18 |
base_model_name_or_path = config.base_model_name_or_path
|
19 |
+
|
20 |
if base_model_name_or_path is None:
|
21 |
raise ValueError("Please specify the base model name or path")
|
22 |
+
|
23 |
pipe = StableDiffusionPipeline.from_pretrained(base_model_name_or_path, torch_dtype=dtype)
|
24 |
pipe.unet = PeftModel.from_pretrained(pipe.unet, unet_sub_dir, adapter_name=adapter_name)
|
25 |
+
|
26 |
if os.path.exists(text_encoder_sub_dir):
|
27 |
+
pipe.text_encoder = PeftModel.from_pretrained(pipe.text_encoder, text_encoder_sub_dir, adapter_name=adapter_name)
|
28 |
+
|
|
|
|
|
29 |
if dtype in (torch.float16, torch.bfloat16):
|
30 |
pipe.unet.half()
|
31 |
pipe.text_encoder.half()
|
32 |
+
|
33 |
return pipe
|
34 |
|
35 |
+
def process_prompt(prompt, tokenizer, text_encoder, max_length=77):
|
36 |
+
tokens = tokenizer(prompt, truncation=False, return_tensors="pt")["input_ids"]
|
37 |
+
chunks = [tokens[:, i:i + max_length] for i in range(0, tokens.shape[1], max_length)]
|
38 |
+
|
39 |
+
with torch.no_grad():
|
40 |
+
embeds = [text_encoder(chunk.to(text_encoder.device))[0] for chunk in chunks]
|
41 |
+
|
42 |
+
return torch.cat(embeds, dim=1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
|
44 |
+
def align_embeddings(prompt_embeds, negative_prompt_embeds):
|
45 |
max_length = max(prompt_embeds.shape[1], negative_prompt_embeds.shape[1])
|
46 |
+
return torch.nn.functional.pad(prompt_embeds, (0, 0, 0, max_length - prompt_embeds.shape[1])), \
|
47 |
+
torch.nn.functional.pad(negative_prompt_embeds, (0, 0, 0, max_length - negative_prompt_embeds.shape[1]))
|
48 |
|
49 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
50 |
model_id_default = "CompVis/stable-diffusion-v1-4"
|
51 |
+
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
|
52 |
|
53 |
+
pipe_default = get_lora_sd_pipeline(ckpt_dir='./lora_logos', base_model_name_or_path=model_id_default, dtype=torch_dtype).to(device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
|
|
|
|
|
55 |
def infer(
|
56 |
+
prompt,
|
57 |
+
negative_prompt,
|
58 |
+
width=512,
|
59 |
+
height=512,
|
60 |
+
num_inference_steps=20,
|
61 |
+
model_id='CompVis/stable-diffusion-v1-4',
|
62 |
+
seed=42,
|
63 |
+
guidance_scale=7.0,
|
64 |
+
lora_scale=0.5
|
65 |
+
):
|
66 |
+
generator = torch.Generator(device).manual_seed(seed)
|
67 |
+
|
68 |
+
print(prompt)
|
69 |
+
print(type(prompt))
|
70 |
|
71 |
+
print(negative_prompt)
|
72 |
+
print(type(negative_prompt))
|
73 |
+
|
74 |
+
if model_id != model_id_default:
|
75 |
+
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch_dtype).to(device)
|
76 |
+
prompt_embeds = process_prompt(prompt, pipe.tokenizer, pipe.text_encoder)
|
77 |
+
negative_prompt_embeds = process_prompt(negative_prompt, pipe.tokenizer, pipe.text_encoder)
|
78 |
+
prompt_embeds, negative_prompt_embeds = align_embeddings(prompt_embeds, negative_prompt_embeds)
|
79 |
+
else:
|
80 |
+
pipe = pipe_default
|
81 |
+
prompt_embeds = process_prompt(prompt, pipe.tokenizer, pipe.text_encoder)
|
82 |
+
negative_prompt_embeds = process_prompt(negative_prompt, pipe.tokenizer, pipe.text_encoder)
|
83 |
+
prompt_embeds, negative_prompt_embeds = align_embeddings(prompt_embeds, negative_prompt_embeds)
|
84 |
+
pipe.fuse_lora(lora_scale=lora_scale)
|
85 |
+
|
86 |
params = {
|
87 |
+
'prompt_embeds': prompt_embeds,
|
88 |
+
'negative_prompt_embeds': negative_prompt_embeds,
|
89 |
'guidance_scale': guidance_scale,
|
90 |
'num_inference_steps': num_inference_steps,
|
91 |
'width': width,
|
92 |
'height': height,
|
93 |
'generator': generator,
|
94 |
}
|
95 |
+
|
96 |
+
return pipe(**params).images[0]
|
97 |
+
|
98 |
+
with gr.Blocks() as demo:
|
99 |
+
with gr.Column():
|
100 |
+
gr.Markdown("# DEMO Text-to-Image")
|
101 |
+
model_id = gr.Textbox(label="Model ID", value=model_id_default)
|
102 |
+
prompt = gr.Textbox(label="Prompt")
|
103 |
+
negative_prompt = gr.Textbox(label="Negative prompt")
|
104 |
+
seed = gr.Number(label="Seed", value=42)
|
105 |
+
guidance_scale = gr.Slider(label="Guidance scale", minimum=0.0, maximum=10.0, value=7.0)
|
106 |
+
lora_scale = gr.Slider(label="LoRA scale", minimum=0.0, maximum=1.0, value=0.5)
|
107 |
+
num_inference_steps = gr.Slider(label="Number of inference steps", minimum=1, maximum=50, value=20)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
109 |
with gr.Accordion("Optional Settings", open=False):
|
110 |
+
width = gr.Slider(label="Width", minimum=256, maximum=1024, value=512, step=32)
|
111 |
+
height = gr.Slider(label="Height", minimum=256, maximum=1024, value=512, step=32)
|
112 |
+
|
113 |
+
run_button = gr.Button("Run")
|
114 |
+
result = gr.Image(label="Result")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
115 |
|
116 |
+
run_button.click(
|
117 |
+
fn=infer,
|
|
|
118 |
inputs=[
|
119 |
+
prompt,
|
120 |
+
negative_prompt,
|
121 |
+
width,
|
122 |
+
height,
|
123 |
+
num_inference_steps,
|
124 |
+
model_id, seed,
|
125 |
+
guidance_scale,
|
126 |
+
lora_scale
|
127 |
+
],
|
128 |
+
outputs=result)
|
|
|
|
|
129 |
|
130 |
if __name__ == "__main__":
|
131 |
+
demo.launch()
|