File size: 12,406 Bytes
7d22699 785814d 7d22699 2f87c44 7d22699 785814d 7d22699 3adda7c 7d22699 785814d 7d22699 785814d 7d22699 785814d 7d22699 785814d 7d22699 7989328 7d22699 3adda7c 7d22699 7baa44d 7d22699 785814d 7d22699 785814d 7d22699 785814d 7d22699 785814d 7d22699 785814d 7baa44d 7d22699 2f87c44 7d22699 7baa44d 2f87c44 7baa44d 7d22699 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 |
import gradio as gr
import torch
from PIL import Image
from torchvision import transforms
from diffusers import StableDiffusionPipeline, StableDiffusionImageVariationPipeline, DiffusionPipeline
import numpy as np
import pandas as pd
import math
from transformers import CLIPTextModel, CLIPTokenizer
# model_id = "stabilityai/stable-diffusion-2-1-base"
# text_model_id = "CompVis/stable-diffusion-v-1-4-original"
# text_model_id = "CompVis/stable-diffusion-v1-4"
text_model_id = "runwayml/stable-diffusion-v1-5"
# text_model_id = "stabilityai/stable-diffusion-2-1-base"
model_id = "lambdalabs/sd-image-variations-diffusers"
clip_model_id = "openai/clip-vit-large-patch14-336"
max_tabs = 10
input_images = [None for i in range(max_tabs)]
input_prompts = [None for i in range(max_tabs)]
embedding_plots = [None for i in range(max_tabs)]
# global embedding_base64s
embedding_base64s = [None for i in range(max_tabs)]
# embedding_base64s = gr.State(value=[None for i in range(max_tabs)])
def image_to_embedding(input_im):
tform = transforms.Compose([
transforms.ToTensor(),
transforms.Resize(
(224, 224),
interpolation=transforms.InterpolationMode.BICUBIC,
antialias=False,
),
transforms.Normalize(
[0.48145466, 0.4578275, 0.40821073],
[0.26862954, 0.26130258, 0.27577711]),
])
inp = tform(input_im).to(device)
dtype = next(pipe.image_encoder.parameters()).dtype
image = inp.tile(1, 1, 1, 1).to(device=device, dtype=dtype)
image_embeddings = pipe.image_encoder(image).image_embeds
image_embeddings = image_embeddings[0]
image_embeddings_np = image_embeddings.cpu().detach().numpy()
return image_embeddings_np
def prompt_to_embedding(prompt):
# inputs = processor(prompt, images=imgs, return_tensors="pt", padding=True)
inputs = processor(prompt, return_tensors="pt", padding='max_length', max_length=77)
# labels = torch.tensor(labels)
# prompt_tokens = inputs.input_ids[0]
prompt_tokens = inputs.input_ids
# image = inputs.pixel_values
with torch.no_grad():
prompt_embededdings = model.get_text_features(prompt_tokens.to(device))
prompt_embededdings = prompt_embededdings[0].cpu().detach().numpy()
return prompt_embededdings
def embedding_to_image(embeddings):
size = math.ceil(math.sqrt(embeddings.shape[0]))
image_embeddings_square = np.pad(embeddings, (0, size**2 - embeddings.shape[0]), 'constant')
image_embeddings_square.resize(size,size)
embedding_image = Image.fromarray(image_embeddings_square, mode="L")
return embedding_image
def embedding_to_base64(embeddings):
import base64
# ensure float16
embeddings = embeddings.astype(np.float16)
embeddings_b64 = base64.urlsafe_b64encode(embeddings).decode()
return embeddings_b64
def base64_to_embedding(embeddings_b64):
import base64
embeddings = base64.urlsafe_b64decode(embeddings_b64)
embeddings = np.frombuffer(embeddings, dtype=np.float16)
# embeddings = torch.tensor(embeddings)
return embeddings
def main(
# input_im,
embeddings,
scale=3.0,
n_samples=4,
steps=25,
seed=None
):
if seed == None:
seed = np.random.randint(2147483647)
# if device contains cuda
if device.type == 'cuda':
generator = torch.Generator(device=device).manual_seed(int(seed))
else:
generator = torch.Generator().manual_seed(int(seed)) # use cpu as does not work on mps
embeddings = base64_to_embedding(embeddings)
embeddings = torch.tensor(embeddings, dtype=torch_size).to(device)
images_list = pipe(
# inp.tile(n_samples, 1, 1, 1),
# [embeddings * n_samples],
embeddings,
guidance_scale=scale,
num_inference_steps=steps,
generator=generator,
)
images = []
for i, image in enumerate(images_list["images"]):
images.append(image)
# images.append(embedding_image)
return images
def on_image_load_update_embeddings(image_data):
# image to embeddings
if image_data is None:
# embeddings = prompt_to_embedding('')
# embeddings_b64 = embedding_to_base64(embeddings)
# return gr.Text.update(embeddings_b64)
return gr.Text.update('')
embeddings = image_to_embedding(image_data)
embeddings_b64 = embedding_to_base64(embeddings)
return gr.Text.update(embeddings_b64)
def on_prompt_change_update_embeddings(prompt):
# prompt to embeddings
if prompt is None or prompt == "":
embeddings = prompt_to_embedding('')
embeddings_b64 = embedding_to_base64(embeddings)
return gr.Text.update(embedding_to_base64(embeddings))
embeddings = prompt_to_embedding(prompt)
embeddings_b64 = embedding_to_base64(embeddings)
return gr.Text.update(embeddings_b64)
# def on_embeddings_changed_update_average_embeddings(last_embedding_base64):
# def on_embeddings_changed_update_average_embeddings(embedding_base64s):
def on_embeddings_changed_update_average_embeddings(embedding_base64s_state, embedding_base64, idx):
# global embedding_base64s
final_embedding = None
num_embeddings = 0
embedding_base64s_state[idx] = embedding_base64 if embedding_base64 != '' else None
# for textbox in embedding_base64s:
# embedding_base64 = textbox.value
for embedding_base64 in embedding_base64s_state:
if embedding_base64 is None or embedding_base64 == "":
continue
embedding = base64_to_embedding(embedding_base64)
if final_embedding is None:
final_embedding = embedding
else:
final_embedding = final_embedding + embedding
num_embeddings += 1
if final_embedding is None:
# embeddings = prompt_to_embedding('')
# embeddings_b64 = embedding_to_base64(embeddings)
# return gr.Text.update(embeddings_b64)
return gr.Text.update('')
final_embedding = final_embedding / num_embeddings
embeddings_b64 = embedding_to_base64(final_embedding)
return gr.Text.update(embeddings_b64)
def on_embeddings_changed_update_plot(embeddings_b64):
# plot new embeddings
if embeddings_b64 is None or embeddings_b64 == "":
data = pd.DataFrame({
'embedding': [],
'index': []})
return gr.LinePlot.update(data,
x="index",
y="embedding",
# color="country",
title="Embeddings",
# stroke_dash="cluster",
# x_lim=[1950, 2010],
tooltip=['index', 'embedding'],
# stroke_dash_legend_title="Country Cluster",
# height=300,
width=0)
embeddings = base64_to_embedding(embeddings_b64)
data = pd.DataFrame({
'embedding': embeddings,
'index': [n for n in range(len(embeddings))]})
return gr.LinePlot.update(data,
x="index",
y="embedding",
# color="country",
title="Embeddings",
# stroke_dash="cluster",
# x_lim=[1950, 2010],
tooltip=['index', 'embedding'],
# stroke_dash_legend_title="Country Cluster",
# height=300,
width=embeddings.shape[0])
device = torch.device("mps" if torch.backends.mps.is_available() else "cuda:0" if torch.cuda.is_available() else "cpu")
torch_size = torch.float16 if device == ('cuda') else torch.float32
# torch_size = torch.float32
pipe = StableDiffusionPipeline.from_pretrained(
model_id,
custom_pipeline="pipeline.py",
torch_dtype=torch_size,
# , revision="fp16",
requires_safety_checker = False, safety_checker=None,
text_encoder = CLIPTextModel,
tokenizer = CLIPTokenizer,
)
pipe = pipe.to(device)
from transformers import AutoProcessor, AutoModel
processor = AutoProcessor.from_pretrained(clip_model_id)
model = AutoModel.from_pretrained(clip_model_id)
model = model.to(device)
examples = [
["frog.png", 3, 1, 25, 0],
["img0.jpg", 3, 1, 25, 0],
["img1.jpg", 3, 1, 25, 0],
["img2.jpg", 3, 1, 25, 0],
["img3.jpg", 3, 1, 25, 0],
]
with gr.Blocks() as demo:
with gr.Row():
gr.Markdown(
"""# Soho-Clip
A tool for exploring CLIP embedding spaces.
My interest is to use CLIP for image/video understanding (see [CLIP_visual-spatial-reasoning](https://github.com/Sohojoe/CLIP_visual-spatial-reasoning).)
Try it out by uploading a few images/add text prompts and generate images of the average of their embeddings
""")
with gr.Row():
for i in range(max_tabs):
with gr.Tab(f"Input {i+1}"):
with gr.Row():
with gr.Column(scale=1, min_width=240):
input_images[i] = gr.Image()
with gr.Column(scale=3, min_width=600):
embedding_plots[i] = gr.LinePlot(show_label=False).style(container=False)
# input_image.change(on_image_load, inputs= [input_image, plot])
with gr.Row():
with gr.Column(scale=2, min_width=240):
input_prompts[i] = gr.Textbox()
with gr.Column(scale=3, min_width=600):
with gr.Accordion(f"Embeddings (base64)", open=False):
embedding_base64s[i] = gr.Textbox(show_label=False)
with gr.Row():
average_embedding_plot = gr.LinePlot(show_label=True, label="Average Embeddings (base64)").style(container=False)
with gr.Row():
with gr.Accordion(f"Avergage embeddings in base 64", open=False):
average_embedding_base64 = gr.Textbox(show_label=False)
with gr.Row():
submit = gr.Button("Generate images")
with gr.Row():
with gr.Column(scale=1, min_width=200):
scale = gr.Slider(0, 25, value=3, step=1, label="Guidance scale")
with gr.Column(scale=1, min_width=200):
n_samples = gr.Slider(1, 4, value=1, step=1, label="Number images")
with gr.Column(scale=1, min_width=200):
steps = gr.Slider(5, 50, value=25, step=5, label="Steps")
with gr.Column(scale=1, min_width=200):
seed = gr.Number(None, label="Seed (blank = random)", precision=0)
with gr.Row():
output = gr.Gallery(label="Generated variations")
embedding_base64s_state = gr.State(value=[None for i in range(max_tabs)])
for i in range(max_tabs):
input_images[i].change(on_image_load_update_embeddings, input_images[i], [embedding_base64s[i]])
input_prompts[i].submit(on_prompt_change_update_embeddings, input_prompts[i], [embedding_base64s[i]])
embedding_base64s[i].change(on_embeddings_changed_update_plot, embedding_base64s[i], [embedding_plots[i]])
# embedding_plots[i].change(on_plot_changed, embedding_base64s[i], average_embedding_base64)
# embedding_plots[i].change(on_embeddings_changed_update_average_embeddings, embedding_base64s[i], average_embedding_base64)
idx_state = gr.State(value=i)
embedding_base64s[i].change(on_embeddings_changed_update_average_embeddings, [embedding_base64s_state, embedding_base64s[i], idx_state], average_embedding_base64)
average_embedding_base64.change(on_embeddings_changed_update_plot, average_embedding_base64, average_embedding_plot)
# submit.click(main, inputs= [embedding_base64s[0], scale, n_samples, steps, seed], outputs=output)
submit.click(main, inputs= [average_embedding_base64, scale, n_samples, steps, seed], outputs=output)
output.style(grid=2)
with gr.Row():
gr.Markdown(
"""### Initial Features
- Combine up to 10 Images and/or text inputs to create an average embedding space.
- View embedding spaces as graph
- Generate a new image based on the average embedding space
### Known limitations
- Text input is a little off (requires fine tuning and I'm having issues with that at the moment)
- It can only generate a single image at a time
- Not easy to use the sample images
### Acknowledgements
- I heavily build on Justin Pinkney's Experiments in Image Variation (see https://www.justinpinkney.com/image-variation-experiments). Please credit them if you use this work.
- [CLIP] (https://openai.com/blog/clip/)
- [Stable Diffusion](https://github.com/CompVis/stable-diffusion)
""")
if __name__ == "__main__":
demo.launch() |