File size: 12,231 Bytes
c25e2cc
3cdacdf
 
 
 
 
6255790
 
5e25b83
6255790
98c6b44
c25e2cc
6255790
 
 
 
 
 
 
 
 
 
 
 
 
e79152d
 
6255790
 
 
09eb804
6255790
 
 
6a2fd41
6255790
 
 
 
 
e79152d
 
6255790
 
 
 
 
5e25b83
d7187bf
5e25b83
9b96547
f55706c
27e096e
6255790
7f61c74
 
 
fd93e8a
 
 
 
 
 
 
 
 
 
 
 
 
7f61c74
b12e6a1
 
 
33f6feb
 
 
dd85adc
33f6feb
 
b12e6a1
 
b13a3d4
 
 
 
 
33f6feb
 
 
b13a3d4
33f6feb
 
b13a3d4
 
652f06c
 
2331708
652f06c
 
33f6feb
 
 
652f06c
33f6feb
 
652f06c
 
7f61c74
 
9cd2450
 
 
017df60
 
 
 
05f89f0
 
 
017df60
05f89f0
 
3be3aae
017df60
9cd2450
26f7dab
9df32ef
6255790
017df60
 
7ff8616
017df60
 
 
6255790
7ff8616
6f62fc3
017df60
6255790
774cc5f
9cd2450
017df60
 
 
 
 
 
 
 
 
 
 
 
3cc82a2
bc09c01
7cbd357
88f076f
98c6b44
88f076f
98c6b44
88f076f
 
93e1172
88f076f
 
 
 
 
98c6b44
 
 
 
 
 
61507ea
98c6b44
 
 
 
 
 
5e25b83
88f076f
 
98c6b44
 
69d2e27
5e25b83
 
 
017df60
6f62fc3
1a248f3
3489b04
017df60
9cd2450
6255790
9156300
 
 
 
 
 
88f076f
 
 
 
db50056
660a4aa
393519b
3489b04
393519b
aefef30
 
 
c635e15
1ff3548
 
3489b04
 
7bb8383
bccbcd8
017df60
 
 
 
 
7bb8383
017df60
 
 
d64d565
77d316c
f948a49
853c8df
d58e1aa
77d316c
 
 
 
12fd402
 
853c8df
 
 
 
 
12fd402
7bb8383
 
9cd2450
b9a2245
9cd2450
7bb8383
b885715
448a301
4b95bff
 
12fd402
4b95bff
017df60
bc09c01
017df60
001613c
017df60
d2cde50
017df60
4b95bff
d0a0320
001613c
 
88f076f
448a301
7bb8383
fdf34ba
7bb8383
9cd2450
017df60
 
 
 
b9a2245
9cd2450
017df60
 
 
 
 
 
 
 
 
9cd2450
017df60
9cd2450
 
 
7078734
7bb8383
017df60
 
 
 
 
 
 
 
 
 
 
 
bc09c01
277aca5
9cd2450
017df60
 
 
 
 
 
277aca5
 
b12e6a1
 
 
7bb0fbf
 
 
 
 
 
 
 
9cd84bd
33f6feb
2b4fe45
539fd6d
 
 
7f61c74
3489b04
 
a93910d
b30a076
3489b04
 
 
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
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
import gradio as gr
import torch
import requests
from io import BytesIO
from diffusers import StableDiffusionPipeline
from diffusers import DDIMScheduler
from utils import *
from inversion_utils import *
from modified_pipeline_semantic_stable_diffusion import SemanticStableDiffusionPipeline
from torch import autocast, inference_mode
import re

def invert(x0, prompt_src="", num_diffusion_steps=100, cfg_scale_src = 3.5, eta = 1):

  #  inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf, 
  #  based on the code in https://github.com/inbarhub/DDPM_inversion
   
  #  returns wt, zs, wts:
  #  wt - inverted latent
  #  wts - intermediate inverted latents
  #  zs - noise maps

  sd_pipe.scheduler.set_timesteps(num_diffusion_steps)

  # vae encode image
  with autocast("cuda"), inference_mode():
      w0 = (sd_pipe.vae.encode(x0).latent_dist.mode() * 0.18215).float()

  # find Zs and wts - forward process
  wt, zs, wts = inversion_forward_process(sd_pipe, w0, etas=eta, prompt=prompt_src, cfg_scale=cfg_scale_src, prog_bar=True, num_inference_steps=num_diffusion_steps)
  return zs, wts



def sample(zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1):

    # reverse process (via Zs and wT)
    w0, _ = inversion_reverse_process(sd_pipe, xT=wts[skip], etas=eta, prompts=[prompt_tar], cfg_scales=[cfg_scale_tar], prog_bar=True, zs=zs[skip:])
    
    # vae decode image
    with autocast("cuda"), inference_mode():
        x0_dec = sd_pipe.vae.decode(1 / 0.18215 * w0).sample
    if x0_dec.dim()<4:
        x0_dec = x0_dec[None,:,:,:]
    img = image_grid(x0_dec)
    return img

# load pipelines
sd_model_id = "runwayml/stable-diffusion-v1-5"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
sd_pipe = StableDiffusionPipeline.from_pretrained(sd_model_id).to(device)
sd_pipe.scheduler = DDIMScheduler.from_config(sd_model_id, subfolder = "scheduler")
sem_pipe = SemanticStableDiffusionPipeline.from_pretrained(sd_model_id).to(device)


def get_example():
    case = [
        [
            'examples/source_a_cat_sitting_next_to_a_mirror.jpeg', 
            'a cat sitting next to a mirror',
            'watercolor painting of a cat sitting next to a mirror',
            100,
            36,
            15,
            '+Schnauzer dog, -cat',
            5.5,
            1,
            'examples/ddpm_watercolor_painting_a_cat_sitting_next_to_a_mirror.png', 
            'examples/ddpm_sega_watercolor_painting_a_cat_sitting_next_to_a_mirror_plus_dog_minus_cat.png'
             ],
        [
            'examples/source_a_man_wearing_a_brown_hoodie_in_a_crowded_street.jpeg', 
            'a man wearing a brown hoodie in a crowded street',
            'a robot wearing a brown hoodie in a crowded street',
            100,
            36,
            15,
            '+painting',
            10,
            1,
            'examples/ddpm_a_robot_wearing_a_brown_hoodie_in_a_crowded_street.png', 
            'examples/ddpm_sega_painting_of_a_robot_wearing_a_brown_hoodie_in_a_crowded_street.png'
             ],
    [
            'examples/source_wall_with_framed_photos.jpeg', 
            '',
            '',
            100,
            36,
            15,
            '+pink drawings of muffins',
            10,
            1,
            'examples/ddpm_wall_with_framed_photos.png', 
            'examples/ddpm_sega_plus_pink_drawings_of_muffins.png'
             ],
    [
            'examples/source_an_empty_room_with_concrete_walls.jpg', 
            'an empty room with concrete walls',
            'glass walls',
            100,
            36,
            17,
            '+giant elephant',
            10,
            1,
            'examples/ddpm_glass_walls.png', 
            'examples/ddpm_sega_glass_walls_gian_elephant.png'
             ]]
    return case


def invert_and_reconstruct(
                    input_image, 
                    do_inversion, 
                    wts, zs, 
                    seed,
                    src_prompt ="", 
                    tar_prompt="", 
                    steps=100,
                    src_cfg_scale = 3.5,
                    skip=36,
                    tar_cfg_scale=15,
                    # neg_guidance=False,
                    
):
    torch.manual_seed(seed)
    x0 = load_512(input_image, device=device)

    if do_inversion:
        # invert and retrieve noise maps and latent
        zs_tensor, wts_tensor = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps, cfg_scale_src=src_cfg_scale)
        wts = gr.State(value=wts_tensor)
        zs = gr.State(value=zs_tensor)
        do_inversion = False

    output = sample(zs.value, wts.value, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale)

    return output, wts, zs, do_inversion


    
def edit(input_image,
            do_inversion, 
            wts, zs, seed,
            src_prompt ="", 
            tar_prompt="", 
            steps=100,
            skip=36,
            tar_cfg_scale=15,
            edit_concept="",
            sega_edit_guidance=10,
            warm_up=None,
            # neg_guidance=False,

   ):
       
    # SEGA
    # parse concepts and neg guidance 
    edit_concepts = edit_concept.split(",")
    num_concepts = len(edit_concepts)
    neg_guidance =[] 
    for edit_concept in edit_concepts:
        edit_concept=edit_concept.strip(" ")
        if edit_concept.startswith("-"):
            neg_guidance.append(True)
        else:
            neg_guidance.append(False)
    edit_concepts = [concept.strip("+|-") for concept in edit_concepts]
                        
    # parse warm-up steps
    default_warm_up_steps = [1]*num_concepts
    if warm_up:
        digit_pattern = re.compile(r"^\d+$")
        warm_up_steps_str = warm_up.split(",")
        for i,num_steps in enumerate(warm_up_steps_str[:num_concepts]):
            if not digit_pattern.match(num_steps):
                raise gr.Error("Invalid value for warm-up steps, using 1 instead")
            else:
                default_warm_up_steps[i] = int(num_steps)
        
        
    editing_args = dict(
    editing_prompt = edit_concepts,
    reverse_editing_direction = neg_guidance,
    edit_warmup_steps=default_warm_up_steps,
    edit_guidance_scale=[sega_edit_guidance]*num_concepts, 
    edit_threshold=[.95]*num_concepts,
    edit_momentum_scale=0.5, 
    edit_mom_beta=0.6 
  )
    latnets = wts.value[skip].expand(1, -1, -1, -1)
    sega_out = sem_pipe(prompt=tar_prompt,eta=1, latents=latnets, guidance_scale = tar_cfg_scale,
                        num_images_per_prompt=1,  
                        num_inference_steps=steps, 
                        use_ddpm=True,  wts=wts.value, zs=zs.value[skip:], **editing_args)
    return sega_out.images[0]


def randomize_seed_fn(seed, randomize_seed):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    return seed

########
# demo #
########
                        
intro = """
<h1 style="font-weight: 1400; text-align: center; margin-bottom: 7px;">
   Edit Friendly DDPM X Semantic Guidance
</h1>
<p style="font-size: 0.9rem; text-align: center; margin: 0rem; line-height: 1.2em; margin-top:1em">
<a href="https://arxiv.org/abs/2301.12247" style="text-decoration: underline;" target="_blank">An Edit Friendly DDPM Noise Space:
Inversion and Manipulations </a> X
<a href="https://arxiv.org/abs/2301.12247" style="text-decoration: underline;" target="_blank">SEGA: Instructing Diffusion using Semantic Dimensions</a>
<p/>
<p style="font-size: 0.9rem; margin: 0rem; line-height: 1.2em; margin-top:1em">
For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
<a href="https://huggingface.co/spaces/LinoyTsaban/ddpm_sega?duplicate=true">
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
<p/>"""
with gr.Blocks(css='style.css') as demo:
    
    def reset_do_inversion():
        do_inversion = True
        return do_inversion

    gr.HTML(intro)
    wts = gr.State()
    zs = gr.State()
    do_inversion = gr.State(value=True)
         
    with gr.Row():
        input_image = gr.Image(label="Input Image", interactive=True)
        ddpm_edited_image = gr.Image(label=f"DDPM Reconstructed Image", interactive=False, visible=False)
        sega_edited_image = gr.Image(label=f"DDPM + SEGA Edited Image", interactive=False)
        input_image.style(height=512, width=512)
        ddpm_edited_image.style(height=512, width=512)
        sega_edited_image.style(height=512, width=512)

    with gr.Row():
        tar_prompt = gr.Textbox(lines=1, label="Target Prompt", interactive=True, placeholder="")
        with gr.Accordion("SEGA Concepts", open=False):
            with gr.Column(scale=1, min_width=100):
                edit_concept = gr.Textbox(lines=1, label="SEGA Edit Concepts", visible = True, interactive=True)
            # with gr.Column(scale=1, min_width=100):
                
         
    with gr.Row():
        with gr.Column(scale=1, min_width=100):
            invert_button = gr.Button("Invert")
        with gr.Column(scale=1, min_width=100):
            edit_button = gr.Button("Edit")

    with gr.Accordion("Advanced Options", open=False):
        with gr.Row():
            with gr.Column():
                #inversion
                src_prompt = gr.Textbox(lines=1, label="Source Prompt", interactive=True, placeholder="")
                steps = gr.Number(value=100, precision=0, label="Num Diffusion Steps", interactive=True)
                src_cfg_scale = gr.Number(value=3.5, label=f"Source Guidance Scale", interactive=True)
                seed = gr.Number(value=0, precision=0, label="Seed", interactive=True)
                randomize_seed = gr.Checkbox(label='Randomize seed', value=True)
            with gr.Column():    
                # reconstruction
                skip = gr.Slider(minimum=0, maximum=40, value=36, label="Skip Steps", interactive=True)
                tar_cfg_scale = gr.Slider(minimum=7, maximum=18,value=15, label=f"Guidance Scale", interactive=True)  
                sega_edit_guidance = gr.Slider(value=10, label=f"SEGA Edit Guidance Scale", interactive=True)
                warm_up = gr.Textbox(label=f"SEGA Warm-up Steps", interactive=True, placeholder="type #warm-up steps for each concpets (e.g. 2,7,5...")

            
            # neg_guidance = gr.Checkbox(label="SEGA Negative Guidance")
          

    # gr.Markdown(help_text)

    invert_button.click(
        fn = randomize_seed_fn,
        inputs = [seed, randomize_seed],
        outputs = [seed]
    ).then(
        fn=invert_and_reconstruct,
        inputs=[input_image, 
                do_inversion, 
                wts, zs, 
                seed,
                src_prompt, 
                tar_prompt, 
                steps,
                src_cfg_scale,
                skip,
                tar_cfg_scale,          
        ],
        outputs=[ddpm_edited_image, wts, zs, do_inversion],
    )

    edit_button.click(
        fn=edit,
        inputs=[input_image, 
                do_inversion, 
                wts, zs, 
                seed,
                src_prompt, 
                tar_prompt, 
                steps,
                skip,
                tar_cfg_scale,
                edit_concept,
                sega_edit_guidance,
                warm_up,
                # neg_guidance,

        ],
        outputs=[sega_edited_image],
        
    )

    input_image.change(
        fn = reset_do_inversion,
        outputs = [do_inversion]
    )

    gr.Examples(
        label='Examples', 
        examples=get_example(), 
        inputs=[input_image, src_prompt, tar_prompt, steps,
                    # src_cfg_scale,
                    skip,
                    tar_cfg_scale,
                    edit_concept,
                    sega_edit_guidance,
                    warm_up,
                    # neg_guidance,
                    ddpm_edited_image, sega_edited_image
               ],
        outputs=[ddpm_edited_image, sega_edited_image],
        # fn=edit,
        # cache_examples=True
    )



demo.queue()
demo.launch(share=False)