File size: 14,273 Bytes
b6dd358
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
85d241c
b6dd358
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b2a8658
 
 
 
 
 
aab7fa9
 
 
 
 
 
 
b6dd358
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de4aa1c
b6dd358
3746f14
 
 
 
 
 
 
 
 
b6dd358
3746f14
 
 
 
 
 
 
 
0f09c7c
3746f14
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f09c7c
3746f14
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f09c7c
3746f14
 
 
 
 
 
 
 
b6dd358
 
 
 
 
3746f14
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b6dd358
 
 
 
dfa70ec
 
 
 
 
 
89023a7
 
 
 
dfa70ec
 
 
 
 
 
 
e660aea
7ee08c3
04dbeac
dfa70ec
 
 
c1ef964
dfa70ec
c1ef964
dfa70ec
c1ef964
3746f14
dfa70ec
 
 
 
 
c1ef964
dfa70ec
 
 
 
c1ef964
dfa70ec
c1ef964
dfa70ec
 
 
3746f14
dfa70ec
 
 
 
 
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
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
# %%

# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto.  Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.

from networks.mat import Generator
import gradio as gr
import gradio.components as gc
import base64
import glob
import os
import random
import re
from http import HTTPStatus
from io import BytesIO
from typing import Dict, List, NamedTuple, Optional, Tuple

import click
import cv2
import numpy as np
import PIL.Image
import torch
import torch.nn.functional as F
from PIL import Image, ImageDraw, ImageOps
from pydantic import BaseModel

import dnnlib
import legacy


pyspng = None


def num_range(s: str) -> List[int]:
    '''Accept either a comma separated list of numbers 'a,b,c' or a range 'a-c' and return as a list of ints.'''

    range_re = re.compile(r'^(\d+)-(\d+)$')
    m = range_re.match(s)
    if m:
        return list(range(int(m.group(1)), int(m.group(2))+1))
    vals = s.split(',')
    return [int(x) for x in vals]


def copy_params_and_buffers(src_module, dst_module, require_all=False):
    assert isinstance(src_module, torch.nn.Module)
    assert isinstance(dst_module, torch.nn.Module)
    src_tensors = {name: tensor for name,
                   tensor in named_params_and_buffers(src_module)}
    for name, tensor in named_params_and_buffers(dst_module):
        assert (name in src_tensors) or (not require_all)
        if name in src_tensors:
            tensor.copy_(src_tensors[name].detach()).requires_grad_(
                tensor.requires_grad)


def params_and_buffers(module):
    assert isinstance(module, torch.nn.Module)
    return list(module.parameters()) + list(module.buffers())


def named_params_and_buffers(module):
    assert isinstance(module, torch.nn.Module)
    return list(module.named_parameters()) + list(module.named_buffers())


class Inpainter:
    def __init__(self,
                 network_pkl,
                 resolution=512,
                 truncation_psi=1,
                 noise_mode='const',
                 sdevice='cpu'
                 ):
        self.resolution = resolution
        self.truncation_psi = truncation_psi
        self.noise_mode = noise_mode
        print(f'Loading networks from: {network_pkl}')
        self.device = torch.device(sdevice)
        with dnnlib.util.open_url(network_pkl) as f:
            G_saved = (
                legacy.load_network_pkl(f)
                ['G_ema']
                .to(self.device)
                .eval()
                .requires_grad_(False))  # type: ignore
        net_res = 512 if resolution > 512 else resolution
        self.G = (
            Generator(
                z_dim=512,
                c_dim=0,
                w_dim=512,
                img_resolution=net_res,
                img_channels=3
            )
            .to(self.device)
            .eval()
            .requires_grad_(False)
        )
        copy_params_and_buffers(G_saved,  self.G, require_all=True)

    def generate_images2(
        self,
        dpath: List[PIL.Image.Image],
        mpath: List[Optional[PIL.Image.Image]],
        seed: int = 42,
    ):
        """
        Generate images using pretrained network pickle.
        """
        resolution = self.resolution
        truncation_psi = self.truncation_psi
        noise_mode = self.noise_mode
        # seed = 240  # pick up a random number

        def seed_all(seed):
            random.seed(seed)
            np.random.seed(seed)
            torch.manual_seed(seed)
            torch.cuda.manual_seed(seed)
        if seed is not None:
            seed_all(seed)

        # no Labels.
        label = torch.zeros([1,  self.G.c_dim], device=self.device)

        def read_image(image):
            image = np.array(image)
            if image.ndim == 2:
                image = image[:, :, np.newaxis]  # HW => HWC
                image = np.repeat(image, 3, axis=2)
            image = image.transpose(2, 0, 1)  # HWC => CHW
            image = image[:3]
            return image
        if resolution != 512:
            noise_mode = 'random'
        results = []
        with torch.no_grad():
            for i, (ipath, m) in enumerate(zip(dpath, mpath)):
                if seed is None:
                    seed_all(i)

                image = read_image(ipath)
                image = (torch.from_numpy(image).float().to(
                    self. device) / 127.5 - 1).unsqueeze(0)

                mask = np.array(m).astype(np.float32) / 255.0
                mask = torch.from_numpy(mask).float().to(
                    self. device).unsqueeze(0).unsqueeze(0)

                z = torch.from_numpy(np.random.randn(
                    1,  self.G.z_dim)).to(self.device)
                output = self.G(image, mask, z, label,
                                truncation_psi=truncation_psi, noise_mode=noise_mode)
                output = (output.permute(0, 2, 3, 1) * 127.5 +
                          127.5).round().clamp(0, 255).to(torch.uint8)
                output = output[0].cpu().numpy()
                results.append(PIL.Image.fromarray(output, 'RGB'))

        return results


# if __name__ == "__main__":
#     generate_images()  # pylint: disable=no-value-for-parameter

# ----------------------------------------------------------------------------
def mask_to_alpha(img, mask):
    img = img.copy()
    img.putalpha(mask)
    return img


def blend(src, target, mask):
    mask = np.expand_dims(mask, axis=-1)
    result = (1-mask) * src + mask * target
    return Image.fromarray(result.astype(np.uint8))


def pad(img, size=(128, 128), tosize=(512, 512), border=1):
    if isinstance(size, float):
        size = (int(img.size[0] * size), int(img.size[1] * size))
    # remove border
    w, h = tosize

    new_img = Image.new('RGBA', (w, h))

    rimg = img.resize(size, resample=Image.Resampling.NEAREST)
    rimg = ImageOps.crop(rimg, border=border)
    tw, th = size
    tw, th = tw - border*2, th - border*2
    tc = ((w-tw)//2, (h-th)//2)

    new_img.paste(rimg, tc)
    mask = Image.new('L', (w, h))
    white = Image.new('L', (tw, th), 255)
    mask.paste(white, tc)

    if 'A' in rimg.getbands():
        mask.paste(rimg.getchannel('A'), tc)
    return new_img, mask


def b64_to_img(b64):
    return Image.open(BytesIO(base64.b64decode(b64)))


def img_to_b64(img):
    with BytesIO() as f:
        img.save(f, format='PNG')
        return base64.b64encode(f.getvalue()).decode('utf-8')


class Predictor:
    def __init__(self):
        """Load the model into memory to make running multiple predictions efficient"""
        self.models = {
            "places2": Inpainter(
                network_pkl='models/Places_512_FullData.pkl',
                resolution=512,
                truncation_psi=1.,
                noise_mode='const',
            ),
            "places2+laion300k": Inpainter(
                network_pkl='models/Places_512_FullData+LAION300k.pkl',
                resolution=512,
                truncation_psi=1.,
                noise_mode='const',
            ),
            "places2+laion300k+laion300k(opmasked)": Inpainter(
                network_pkl='models/Places_512_FullData+LAION300k+OPM300k.pkl',
                resolution=512,
                truncation_psi=1.,
                noise_mode='const',
            ),
            "places2+laion300k+laion1200k(opmasked)": Inpainter(
                network_pkl='models/Places_512_FullData+LAION300k+OPM1200k.pkl',
                resolution=512,
                truncation_psi=1.,
                noise_mode='const',
            ),

        }

    # The arguments and types the model takes as input

    def predict(
        self,
        img: Image.Image,
        tosize=(512, 512),
        border=5,
        seed=42,
        size=0.5,
        model='places2',
    ) -> Image:
        i, m = pad(
            img,
            size=size,  # (328, 328),
            tosize=tosize,
            border=border
        )
        """Run a single prediction on the model"""
        imgs = self.models[model].generate_images2(
            dpath=[i.resize((512, 512), resample=Image.Resampling.NEAREST)],
            mpath=[m.resize((512, 512), resample=Image.Resampling.NEAREST)],
            seed=seed,
        )
        img_op_raw = imgs[0].convert('RGBA')
        img_op_raw = img_op_raw.resize(
            tosize, resample=Image.Resampling.NEAREST)
        inpainted = img_op_raw.copy()

        # paste original image to remove inpainting/scaling artifacts
        inpainted = blend(
            i,
            inpainted,
            1-(np.array(m) / 255)
        )
        minpainted = mask_to_alpha(inpainted, m)
        return inpainted, minpainted,  ImageOps.invert(m)

    def predict_tiled(
        self,
        img: Image.Image,
        tosize=(512, 512),
        border=5,
        seed=42,
        size=0.5,
        model='places2',
    ) -> Image:

        i, morig = pad(
            img,
            size=size,  # (328, 328),
            tosize=tosize,
            border=border
        )
        i.putalpha(morig)
        img = i
        # img.save('0.png')
        assert img.width == img.height
        assert img.width > 512 and img.width < 512*2

        def tile_coords(image, n=2, tile_size=512):
            assert image.width == image.height
            offsets = np.linspace(0, image.width - tile_size, n).astype(int)
            for i in range(n):
                for j in range(n):
                    left = offsets[j]
                    upper = offsets[i]
                    right = left + tile_size
                    lower = upper + tile_size
                    # tile = image.crop((left, upper, right, lower))
                    yield [left, upper, right, lower]

        for ix, tc in enumerate(tile_coords(img, n=2)):
            i = img.crop(tc)
            # i.save(f't{ix}.png')
            m = i.getchannel('A')

            """Run a single prediction on the model"""
            imgs = self.models[model].generate_images2(
                dpath=[i.resize((512, 512), resample=Image.Resampling.NEAREST)],
                mpath=[m.resize((512, 512), resample=Image.Resampling.NEAREST)],
                seed=seed,
            )
            img_op_raw = imgs[0].convert('RGBA')
            # img_op_raw = img_op_raw.resize(tosize, resample=Image.Resampling.NEAREST)
            inpainted = img_op_raw.copy()

            # paste original image to remove inpainting/scaling artifacts
            inpainted = blend(
                i,
                inpainted,
                1-(np.array(m) / 255)
            )
            # inpainted.save(f't{ix}_op.png')
            minpainted = mask_to_alpha(inpainted, m)
            # continue with partially inpainted image
            # since the tiles overlap, the next tile will contain (possibly inpainted) parts of the previous tile
            img.paste(inpainted, tc)

        # restore original alpha channel
        img.putalpha(morig)
        return img.convert('RGB'), img,  ImageOps.invert(img.getchannel('A'))
predictor = Predictor()

# %%


def _outpaint(img, tosize, border, seed, size, model, tiled):
    if tiled:
        img_op = predictor.predict_tiled(
            img,
            border=border,
            seed=seed,
            tosize=(tosize, tosize),
            size=float(size),
            model=model,
        )
    else:
        img_op = predictor.predict(
            img,
            border=border,
            seed=seed,
            tosize=(tosize, tosize),
            size=float(size),
            model=model,
        )
    return img_op
# %%


with gr.Blocks() as demo:
    maturl = 'https://github.com/fenglinglwb/MAT'
    gr.Markdown(f'''
        # MAT Primer for Stable Diffusion
        ## based on MAT: Mask-Aware Transformer for Large Hole Image Inpainting
        ### create a primer for use in stable diffusion outpainting

        i have added 2 example scripts to the repo:
        - outpainting_example1.py  using the inpainting pipeline
        - outpainting_example2.py  using the img2img pipeline. this is basically what i used for the examples below
        ''')

    gr.HTML(f'''<a href="{maturl}">{maturl}</a>''')
    with gr.Box():
        with gr.Row():
            gr.Markdown(f"""example with strength 0.5""")
        with gr.Row():
            gr.HTML("<img src='file/hild.gif'> ")
            gr.HTML("<img src='file/process.gif'>")
            gr.HTML("<img src='file/flagscapes.gif'>")
    btn = gr.Button("Run", variant="primary")
    with gr.Row():
        with gr.Column():
            searchimage = gc.Image(label="image", type='pil', image_mode='RGBA')
            to_size = gc.Slider(1, 1920, 512, step=1, label='output size')
            border = gc.Slider(1, 50, 0, step=1, label='border to crop from the image before outpainting')
            seed = gc.Slider(1, 65536, 10, step=1, label='seed')
            size = gc.Slider(0, 1, .5, step=0.01,label='scale of the image before outpainting')
            tiled = gc.Checkbox(label='tiled: run the network with 4 tiles of size 512x512 . only usable if output size >512 and <1024', value=False)

            model = gc.Dropdown(
                choices=['places2',
                         'places2+laion300k',
                         'places2+laion300k+laion300k(opmasked)',
                         'places2+laion300k+laion1200k(opmasked)'],
                value='places2+laion300k+laion1200k(opmasked)',
                label='model',
            )
        with gr.Column():
            outwithoutalpha = gc.Image(label="primed image without alpha channel", type='pil', image_mode='RGBA')
            mask = gc.Image(label="outpainting mask", type='pil')
            out = gc.Image(label="primed image with alpha channel",type='pil', image_mode='RGBA')

    btn.click(
        fn=_outpaint,
        inputs=[searchimage, to_size, border, seed, size, model,tiled],
        outputs=[outwithoutalpha, out,  mask])


# %% launch
demo.launch()