Spaces:
No application file
No application file
Delete app.py
Browse files
app.py
DELETED
@@ -1,492 +0,0 @@
|
|
1 |
-
# Copyright 2022 Lunar Ring. All rights reserved.
|
2 |
-
# Written by Johannes Stelzer, email stelzer@lunar-ring.ai twitter @j_stelzer
|
3 |
-
#
|
4 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
-
# you may not use this file except in compliance with the License.
|
6 |
-
# You may obtain a copy of the License at
|
7 |
-
#
|
8 |
-
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
-
#
|
10 |
-
# Unless required by applicable law or agreed to in writing, software
|
11 |
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
-
# See the License for the specific language governing permissions and
|
14 |
-
# limitations under the License.
|
15 |
-
|
16 |
-
import os
|
17 |
-
import torch
|
18 |
-
torch.backends.cudnn.benchmark = False
|
19 |
-
torch.set_grad_enabled(False)
|
20 |
-
import numpy as np
|
21 |
-
import warnings
|
22 |
-
warnings.filterwarnings('ignore')
|
23 |
-
import warnings
|
24 |
-
from tqdm.auto import tqdm
|
25 |
-
from PIL import Image
|
26 |
-
from movie_util import MovieSaver, concatenate_movies
|
27 |
-
from latent_blending import LatentBlending
|
28 |
-
from stable_diffusion_holder import StableDiffusionHolder
|
29 |
-
import gradio as gr
|
30 |
-
from dotenv import find_dotenv, load_dotenv
|
31 |
-
import shutil
|
32 |
-
import random
|
33 |
-
from utils import get_time, add_frames_linear_interp
|
34 |
-
from huggingface_hub import hf_hub_download
|
35 |
-
|
36 |
-
|
37 |
-
class BlendingFrontend():
|
38 |
-
def __init__(
|
39 |
-
self,
|
40 |
-
sdh,
|
41 |
-
share=False):
|
42 |
-
r"""
|
43 |
-
Gradio Helper Class to collect UI data and start latent blending.
|
44 |
-
Args:
|
45 |
-
sdh:
|
46 |
-
StableDiffusionHolder
|
47 |
-
share: bool
|
48 |
-
Set true to get a shareable gradio link (e.g. for running a remote server)
|
49 |
-
"""
|
50 |
-
self.share = share
|
51 |
-
|
52 |
-
# UI Defaults
|
53 |
-
self.num_inference_steps = 30
|
54 |
-
self.depth_strength = 0.25
|
55 |
-
self.seed1 = 420
|
56 |
-
self.seed2 = 420
|
57 |
-
self.prompt1 = ""
|
58 |
-
self.prompt2 = ""
|
59 |
-
self.negative_prompt = ""
|
60 |
-
self.fps = 30
|
61 |
-
self.duration_video = 8
|
62 |
-
self.t_compute_max_allowed = 10
|
63 |
-
|
64 |
-
self.lb = LatentBlending(sdh)
|
65 |
-
self.lb.sdh.num_inference_steps = self.num_inference_steps
|
66 |
-
self.init_parameters_from_lb()
|
67 |
-
self.init_save_dir()
|
68 |
-
|
69 |
-
# Vars
|
70 |
-
self.list_fp_imgs_current = []
|
71 |
-
self.recycle_img1 = False
|
72 |
-
self.recycle_img2 = False
|
73 |
-
self.list_all_segments = []
|
74 |
-
self.dp_session = ""
|
75 |
-
self.user_id = None
|
76 |
-
|
77 |
-
def init_parameters_from_lb(self):
|
78 |
-
r"""
|
79 |
-
Automatically init parameters from latentblending instance
|
80 |
-
"""
|
81 |
-
self.height = self.lb.sdh.height
|
82 |
-
self.width = self.lb.sdh.width
|
83 |
-
self.guidance_scale = self.lb.guidance_scale
|
84 |
-
self.guidance_scale_mid_damper = self.lb.guidance_scale_mid_damper
|
85 |
-
self.mid_compression_scaler = self.lb.mid_compression_scaler
|
86 |
-
self.branch1_crossfeed_power = self.lb.branch1_crossfeed_power
|
87 |
-
self.branch1_crossfeed_range = self.lb.branch1_crossfeed_range
|
88 |
-
self.branch1_crossfeed_decay = self.lb.branch1_crossfeed_decay
|
89 |
-
self.parental_crossfeed_power = self.lb.parental_crossfeed_power
|
90 |
-
self.parental_crossfeed_range = self.lb.parental_crossfeed_range
|
91 |
-
self.parental_crossfeed_power_decay = self.lb.parental_crossfeed_power_decay
|
92 |
-
|
93 |
-
def init_save_dir(self):
|
94 |
-
r"""
|
95 |
-
Initializes the directory where stuff is being saved.
|
96 |
-
You can specify this directory in a ".env" file in your latentblending root, setting
|
97 |
-
DIR_OUT='/path/to/saving'
|
98 |
-
"""
|
99 |
-
load_dotenv(find_dotenv(), verbose=False)
|
100 |
-
self.dp_out = os.getenv("DIR_OUT")
|
101 |
-
if self.dp_out is None:
|
102 |
-
self.dp_out = ""
|
103 |
-
self.dp_imgs = os.path.join(self.dp_out, "imgs")
|
104 |
-
os.makedirs(self.dp_imgs, exist_ok=True)
|
105 |
-
self.dp_movies = os.path.join(self.dp_out, "movies")
|
106 |
-
os.makedirs(self.dp_movies, exist_ok=True)
|
107 |
-
self.save_empty_image()
|
108 |
-
|
109 |
-
def save_empty_image(self):
|
110 |
-
r"""
|
111 |
-
Saves an empty/black dummy image.
|
112 |
-
"""
|
113 |
-
self.fp_img_empty = os.path.join(self.dp_imgs, 'empty.jpg')
|
114 |
-
Image.fromarray(np.zeros((self.height, self.width, 3), dtype=np.uint8)).save(self.fp_img_empty, quality=5)
|
115 |
-
|
116 |
-
def randomize_seed1(self):
|
117 |
-
r"""
|
118 |
-
Randomizes the first seed
|
119 |
-
"""
|
120 |
-
seed = np.random.randint(0, 10000000)
|
121 |
-
self.seed1 = int(seed)
|
122 |
-
print(f"randomize_seed1: new seed = {self.seed1}")
|
123 |
-
return seed
|
124 |
-
|
125 |
-
def randomize_seed2(self):
|
126 |
-
r"""
|
127 |
-
Randomizes the second seed
|
128 |
-
"""
|
129 |
-
seed = np.random.randint(0, 10000000)
|
130 |
-
self.seed2 = int(seed)
|
131 |
-
print(f"randomize_seed2: new seed = {self.seed2}")
|
132 |
-
return seed
|
133 |
-
|
134 |
-
def setup_lb(self, list_ui_vals):
|
135 |
-
r"""
|
136 |
-
Sets all parameters from the UI. Since gradio does not support to pass dictionaries,
|
137 |
-
we have to instead pass keys (list_ui_keys, global) and values (list_ui_vals)
|
138 |
-
"""
|
139 |
-
# Collect latent blending variables
|
140 |
-
self.lb.set_width(list_ui_vals[list_ui_keys.index('width')])
|
141 |
-
self.lb.set_height(list_ui_vals[list_ui_keys.index('height')])
|
142 |
-
self.lb.set_prompt1(list_ui_vals[list_ui_keys.index('prompt1')])
|
143 |
-
self.lb.set_prompt2(list_ui_vals[list_ui_keys.index('prompt2')])
|
144 |
-
self.lb.set_negative_prompt(list_ui_vals[list_ui_keys.index('negative_prompt')])
|
145 |
-
self.lb.guidance_scale = list_ui_vals[list_ui_keys.index('guidance_scale')]
|
146 |
-
self.lb.guidance_scale_mid_damper = list_ui_vals[list_ui_keys.index('guidance_scale_mid_damper')]
|
147 |
-
self.t_compute_max_allowed = list_ui_vals[list_ui_keys.index('duration_compute')]
|
148 |
-
self.lb.num_inference_steps = list_ui_vals[list_ui_keys.index('num_inference_steps')]
|
149 |
-
self.lb.sdh.num_inference_steps = list_ui_vals[list_ui_keys.index('num_inference_steps')]
|
150 |
-
self.duration_video = list_ui_vals[list_ui_keys.index('duration_video')]
|
151 |
-
self.lb.seed1 = list_ui_vals[list_ui_keys.index('seed1')]
|
152 |
-
self.lb.seed2 = list_ui_vals[list_ui_keys.index('seed2')]
|
153 |
-
self.lb.branch1_crossfeed_power = list_ui_vals[list_ui_keys.index('branch1_crossfeed_power')]
|
154 |
-
self.lb.branch1_crossfeed_range = list_ui_vals[list_ui_keys.index('branch1_crossfeed_range')]
|
155 |
-
self.lb.branch1_crossfeed_decay = list_ui_vals[list_ui_keys.index('branch1_crossfeed_decay')]
|
156 |
-
self.lb.parental_crossfeed_power = list_ui_vals[list_ui_keys.index('parental_crossfeed_power')]
|
157 |
-
self.lb.parental_crossfeed_range = list_ui_vals[list_ui_keys.index('parental_crossfeed_range')]
|
158 |
-
self.lb.parental_crossfeed_power_decay = list_ui_vals[list_ui_keys.index('parental_crossfeed_power_decay')]
|
159 |
-
self.num_inference_steps = list_ui_vals[list_ui_keys.index('num_inference_steps')]
|
160 |
-
self.depth_strength = list_ui_vals[list_ui_keys.index('depth_strength')]
|
161 |
-
|
162 |
-
if len(list_ui_vals[list_ui_keys.index('user_id')]) > 1:
|
163 |
-
self.user_id = list_ui_vals[list_ui_keys.index('user_id')]
|
164 |
-
else:
|
165 |
-
# generate new user id
|
166 |
-
self.user_id = ''.join((random.choice('ABCDEFGHIJKLMNOPQRSTUVWXYZ') for i in range(8)))
|
167 |
-
print(f"made new user_id: {self.user_id} at {get_time('second')}")
|
168 |
-
|
169 |
-
def save_latents(self, fp_latents, list_latents):
|
170 |
-
r"""
|
171 |
-
Saves a latent trajectory on disk, in npy format.
|
172 |
-
"""
|
173 |
-
list_latents_cpu = [l.cpu().numpy() for l in list_latents]
|
174 |
-
np.save(fp_latents, list_latents_cpu)
|
175 |
-
|
176 |
-
def load_latents(self, fp_latents):
|
177 |
-
r"""
|
178 |
-
Loads a latent trajectory from disk, converts to torch tensor.
|
179 |
-
"""
|
180 |
-
list_latents_cpu = np.load(fp_latents)
|
181 |
-
list_latents = [torch.from_numpy(l).to(self.lb.device) for l in list_latents_cpu]
|
182 |
-
return list_latents
|
183 |
-
|
184 |
-
def compute_img1(self, *args):
|
185 |
-
r"""
|
186 |
-
Computes the first transition image and returns it for display.
|
187 |
-
Sets all other transition images and last image to empty (as they are obsolete with this operation)
|
188 |
-
"""
|
189 |
-
list_ui_vals = args
|
190 |
-
self.setup_lb(list_ui_vals)
|
191 |
-
fp_img1 = os.path.join(self.dp_imgs, f"img1_{self.user_id}")
|
192 |
-
img1 = Image.fromarray(self.lb.compute_latents1(return_image=True))
|
193 |
-
img1.save(fp_img1 + ".jpg")
|
194 |
-
self.save_latents(fp_img1 + ".npy", self.lb.tree_latents[0])
|
195 |
-
self.recycle_img1 = True
|
196 |
-
self.recycle_img2 = False
|
197 |
-
return [fp_img1 + ".jpg", self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.user_id]
|
198 |
-
|
199 |
-
def compute_img2(self, *args):
|
200 |
-
r"""
|
201 |
-
Computes the last transition image and returns it for display.
|
202 |
-
Sets all other transition images to empty (as they are obsolete with this operation)
|
203 |
-
"""
|
204 |
-
if not os.path.isfile(os.path.join(self.dp_imgs, f"img1_{self.user_id}.jpg")): # don't do anything
|
205 |
-
return [self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.user_id]
|
206 |
-
list_ui_vals = args
|
207 |
-
self.setup_lb(list_ui_vals)
|
208 |
-
|
209 |
-
self.lb.tree_latents[0] = self.load_latents(os.path.join(self.dp_imgs, f"img1_{self.user_id}.npy"))
|
210 |
-
fp_img2 = os.path.join(self.dp_imgs, f"img2_{self.user_id}")
|
211 |
-
img2 = Image.fromarray(self.lb.compute_latents2(return_image=True))
|
212 |
-
img2.save(fp_img2 + '.jpg')
|
213 |
-
self.save_latents(fp_img2 + ".npy", self.lb.tree_latents[-1])
|
214 |
-
self.recycle_img2 = True
|
215 |
-
# fixme save seeds. change filenames?
|
216 |
-
return [self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, fp_img2 + ".jpg", self.user_id]
|
217 |
-
|
218 |
-
def compute_transition(self, *args):
|
219 |
-
r"""
|
220 |
-
Computes transition images and movie.
|
221 |
-
"""
|
222 |
-
list_ui_vals = args
|
223 |
-
self.setup_lb(list_ui_vals)
|
224 |
-
print("STARTING TRANSITION...")
|
225 |
-
fixed_seeds = [self.seed1, self.seed2]
|
226 |
-
# Inject loaded latents (other user interference)
|
227 |
-
self.lb.tree_latents[0] = self.load_latents(os.path.join(self.dp_imgs, f"img1_{self.user_id}.npy"))
|
228 |
-
self.lb.tree_latents[-1] = self.load_latents(os.path.join(self.dp_imgs, f"img2_{self.user_id}.npy"))
|
229 |
-
imgs_transition = self.lb.run_transition(
|
230 |
-
recycle_img1=self.recycle_img1,
|
231 |
-
recycle_img2=self.recycle_img2,
|
232 |
-
num_inference_steps=self.num_inference_steps,
|
233 |
-
depth_strength=self.depth_strength,
|
234 |
-
t_compute_max_allowed=self.t_compute_max_allowed,
|
235 |
-
fixed_seeds=fixed_seeds)
|
236 |
-
print(f"Latent Blending pass finished ({get_time('second')}). Resulted in {len(imgs_transition)} images")
|
237 |
-
|
238 |
-
# Subselect three preview images
|
239 |
-
idx_img_prev = np.round(np.linspace(0, len(imgs_transition) - 1, 5)[1:-1]).astype(np.int32)
|
240 |
-
|
241 |
-
list_imgs_preview = []
|
242 |
-
for j in idx_img_prev:
|
243 |
-
list_imgs_preview.append(Image.fromarray(imgs_transition[j]))
|
244 |
-
|
245 |
-
# Save the preview imgs as jpgs on disk so we are not sending umcompressed data around
|
246 |
-
current_timestamp = get_time('second')
|
247 |
-
self.list_fp_imgs_current = []
|
248 |
-
for i in range(len(list_imgs_preview)):
|
249 |
-
fp_img = os.path.join(self.dp_imgs, f"img_preview_{i}_{current_timestamp}.jpg")
|
250 |
-
list_imgs_preview[i].save(fp_img)
|
251 |
-
self.list_fp_imgs_current.append(fp_img)
|
252 |
-
# Insert cheap frames for the movie
|
253 |
-
imgs_transition_ext = add_frames_linear_interp(imgs_transition, self.duration_video, self.fps)
|
254 |
-
|
255 |
-
# Save as movie
|
256 |
-
self.fp_movie = self.get_fp_video_last()
|
257 |
-
if os.path.isfile(self.fp_movie):
|
258 |
-
os.remove(self.fp_movie)
|
259 |
-
ms = MovieSaver(self.fp_movie, fps=self.fps)
|
260 |
-
for img in tqdm(imgs_transition_ext):
|
261 |
-
ms.write_frame(img)
|
262 |
-
ms.finalize()
|
263 |
-
print("DONE SAVING MOVIE! SENDING BACK...")
|
264 |
-
|
265 |
-
# Assemble Output, updating the preview images and le movie
|
266 |
-
list_return = self.list_fp_imgs_current + [self.fp_movie]
|
267 |
-
return list_return
|
268 |
-
|
269 |
-
def stack_forward(self, prompt2, seed2):
|
270 |
-
r"""
|
271 |
-
Allows to generate multi-segment movies. Sets last image -> first image with all
|
272 |
-
relevant parameters.
|
273 |
-
"""
|
274 |
-
# Save preview images, prompts and seeds into dictionary for stacking
|
275 |
-
if len(self.list_all_segments) == 0:
|
276 |
-
timestamp_session = get_time('second')
|
277 |
-
self.dp_session = os.path.join(self.dp_out, f"session_{timestamp_session}")
|
278 |
-
os.makedirs(self.dp_session)
|
279 |
-
|
280 |
-
idx_segment = len(self.list_all_segments)
|
281 |
-
dp_segment = os.path.join(self.dp_session, f"segment_{str(idx_segment).zfill(3)}")
|
282 |
-
|
283 |
-
self.list_all_segments.append(dp_segment)
|
284 |
-
self.lb.write_imgs_transition(dp_segment)
|
285 |
-
|
286 |
-
fp_movie_last = self.get_fp_video_last()
|
287 |
-
fp_movie_next = self.get_fp_video_next()
|
288 |
-
|
289 |
-
shutil.copyfile(fp_movie_last, fp_movie_next)
|
290 |
-
|
291 |
-
self.lb.tree_latents[0] = self.load_latents(os.path.join(self.dp_imgs, f"img1_{self.user_id}.npy"))
|
292 |
-
self.lb.tree_latents[-1] = self.load_latents(os.path.join(self.dp_imgs, f"img2_{self.user_id}.npy"))
|
293 |
-
self.lb.swap_forward()
|
294 |
-
|
295 |
-
shutil.copyfile(os.path.join(self.dp_imgs, f"img2_{self.user_id}.npy"), os.path.join(self.dp_imgs, f"img1_{self.user_id}.npy"))
|
296 |
-
fp_multi = self.multi_concat()
|
297 |
-
list_out = [fp_multi]
|
298 |
-
|
299 |
-
list_out.extend([os.path.join(self.dp_imgs, f"img2_{self.user_id}.jpg")])
|
300 |
-
list_out.extend([self.fp_img_empty] * 4)
|
301 |
-
list_out.append(gr.update(interactive=False, value=prompt2))
|
302 |
-
list_out.append(gr.update(interactive=False, value=seed2))
|
303 |
-
list_out.append("")
|
304 |
-
list_out.append(np.random.randint(0, 10000000))
|
305 |
-
print(f"stack_forward: fp_multi {fp_multi}")
|
306 |
-
return list_out
|
307 |
-
|
308 |
-
def multi_concat(self):
|
309 |
-
r"""
|
310 |
-
Concatentates all stacked segments into one long movie.
|
311 |
-
"""
|
312 |
-
list_fp_movies = self.get_fp_video_all()
|
313 |
-
# Concatenate movies and save
|
314 |
-
fp_final = os.path.join(self.dp_session, f"concat_{self.user_id}.mp4")
|
315 |
-
concatenate_movies(fp_final, list_fp_movies)
|
316 |
-
return fp_final
|
317 |
-
|
318 |
-
def get_fp_video_all(self):
|
319 |
-
r"""
|
320 |
-
Collects all stacked movie segments.
|
321 |
-
"""
|
322 |
-
list_all = os.listdir(self.dp_movies)
|
323 |
-
str_beg = f"movie_{self.user_id}_"
|
324 |
-
list_user = [l for l in list_all if str_beg in l]
|
325 |
-
list_user.sort()
|
326 |
-
list_user = [os.path.join(self.dp_movies, l) for l in list_user]
|
327 |
-
return list_user
|
328 |
-
|
329 |
-
def get_fp_video_next(self):
|
330 |
-
r"""
|
331 |
-
Gets the filepath of the next movie segment.
|
332 |
-
"""
|
333 |
-
list_videos = self.get_fp_video_all()
|
334 |
-
if len(list_videos) == 0:
|
335 |
-
idx_next = 0
|
336 |
-
else:
|
337 |
-
idx_next = len(list_videos)
|
338 |
-
fp_video_next = os.path.join(self.dp_movies, f"movie_{self.user_id}_{str(idx_next).zfill(3)}.mp4")
|
339 |
-
return fp_video_next
|
340 |
-
|
341 |
-
def get_fp_video_last(self):
|
342 |
-
r"""
|
343 |
-
Gets the current video that was saved.
|
344 |
-
"""
|
345 |
-
fp_video_last = os.path.join(self.dp_movies, f"last_{self.user_id}.mp4")
|
346 |
-
return fp_video_last
|
347 |
-
|
348 |
-
|
349 |
-
if __name__ == "__main__":
|
350 |
-
fp_ckpt = hf_hub_download(repo_id="stabilityai/stable-diffusion-2-1-base", filename="v2-1_512-ema-pruned.ckpt")
|
351 |
-
# fp_ckpt = hf_hub_download(repo_id="stabilityai/stable-diffusion-2-1", filename="v2-1_768-ema-pruned.ckpt")
|
352 |
-
bf = BlendingFrontend(StableDiffusionHolder(fp_ckpt))
|
353 |
-
# self = BlendingFrontend(None)
|
354 |
-
|
355 |
-
with gr.Blocks() as demo:
|
356 |
-
with gr.Row():
|
357 |
-
prompt1 = gr.Textbox(label="prompt 1")
|
358 |
-
prompt2 = gr.Textbox(label="prompt 2")
|
359 |
-
|
360 |
-
with gr.Row():
|
361 |
-
duration_compute = gr.Slider(5, 200, bf.t_compute_max_allowed, step=1, label='compute budget', interactive=True)
|
362 |
-
duration_video = gr.Slider(1, 100, bf.duration_video, step=0.1, label='video duration', interactive=True)
|
363 |
-
height = gr.Slider(256, 2048, bf.height, step=128, label='height', interactive=True)
|
364 |
-
width = gr.Slider(256, 2048, bf.width, step=128, label='width', interactive=True)
|
365 |
-
|
366 |
-
with gr.Accordion("Advanced Settings (click to expand)", open=False):
|
367 |
-
|
368 |
-
with gr.Accordion("Diffusion settings", open=True):
|
369 |
-
with gr.Row():
|
370 |
-
num_inference_steps = gr.Slider(5, 100, bf.num_inference_steps, step=1, label='num_inference_steps', interactive=True)
|
371 |
-
guidance_scale = gr.Slider(1, 25, bf.guidance_scale, step=0.1, label='guidance_scale', interactive=True)
|
372 |
-
negative_prompt = gr.Textbox(label="negative prompt")
|
373 |
-
|
374 |
-
with gr.Accordion("Seed control: adjust seeds for first and last images", open=True):
|
375 |
-
with gr.Row():
|
376 |
-
b_newseed1 = gr.Button("randomize seed 1", variant='secondary')
|
377 |
-
seed1 = gr.Number(bf.seed1, label="seed 1", interactive=True)
|
378 |
-
seed2 = gr.Number(bf.seed2, label="seed 2", interactive=True)
|
379 |
-
b_newseed2 = gr.Button("randomize seed 2", variant='secondary')
|
380 |
-
|
381 |
-
with gr.Accordion("Last image crossfeeding.", open=True):
|
382 |
-
with gr.Row():
|
383 |
-
branch1_crossfeed_power = gr.Slider(0.0, 1.0, bf.branch1_crossfeed_power, step=0.01, label='branch1 crossfeed power', interactive=True)
|
384 |
-
branch1_crossfeed_range = gr.Slider(0.0, 1.0, bf.branch1_crossfeed_range, step=0.01, label='branch1 crossfeed range', interactive=True)
|
385 |
-
branch1_crossfeed_decay = gr.Slider(0.0, 1.0, bf.branch1_crossfeed_decay, step=0.01, label='branch1 crossfeed decay', interactive=True)
|
386 |
-
|
387 |
-
with gr.Accordion("Transition settings", open=True):
|
388 |
-
with gr.Row():
|
389 |
-
parental_crossfeed_power = gr.Slider(0.0, 1.0, bf.parental_crossfeed_power, step=0.01, label='parental crossfeed power', interactive=True)
|
390 |
-
parental_crossfeed_range = gr.Slider(0.0, 1.0, bf.parental_crossfeed_range, step=0.01, label='parental crossfeed range', interactive=True)
|
391 |
-
parental_crossfeed_power_decay = gr.Slider(0.0, 1.0, bf.parental_crossfeed_power_decay, step=0.01, label='parental crossfeed decay', interactive=True)
|
392 |
-
with gr.Row():
|
393 |
-
depth_strength = gr.Slider(0.01, 0.99, bf.depth_strength, step=0.01, label='depth_strength', interactive=True)
|
394 |
-
guidance_scale_mid_damper = gr.Slider(0.01, 2.0, bf.guidance_scale_mid_damper, step=0.01, label='guidance_scale_mid_damper', interactive=True)
|
395 |
-
|
396 |
-
with gr.Row():
|
397 |
-
b_compute1 = gr.Button('compute first image', variant='primary')
|
398 |
-
b_compute_transition = gr.Button('compute transition', variant='primary')
|
399 |
-
b_compute2 = gr.Button('compute last image', variant='primary')
|
400 |
-
|
401 |
-
with gr.Row():
|
402 |
-
img1 = gr.Image(label="1/5")
|
403 |
-
img2 = gr.Image(label="2/5", show_progress=False)
|
404 |
-
img3 = gr.Image(label="3/5", show_progress=False)
|
405 |
-
img4 = gr.Image(label="4/5", show_progress=False)
|
406 |
-
img5 = gr.Image(label="5/5")
|
407 |
-
|
408 |
-
with gr.Row():
|
409 |
-
vid_single = gr.Video(label="current single trans")
|
410 |
-
vid_multi = gr.Video(label="concatented multi trans")
|
411 |
-
|
412 |
-
with gr.Row():
|
413 |
-
b_stackforward = gr.Button('append last movie segment (left) to multi movie (right)', variant='primary')
|
414 |
-
|
415 |
-
with gr.Row():
|
416 |
-
gr.Markdown(
|
417 |
-
"""
|
418 |
-
# Parameters
|
419 |
-
## Main
|
420 |
-
- compute budget: set your waiting time for the transition. high values = better quality
|
421 |
-
- video duration: seconds per segment
|
422 |
-
- height/width: in pixels
|
423 |
-
|
424 |
-
## Diffusion settings
|
425 |
-
- num_inference_steps: number of diffusion steps
|
426 |
-
- guidance_scale: latent blending seems to prefer lower values here
|
427 |
-
- negative prompt: enter negative prompt here, applied for all images
|
428 |
-
|
429 |
-
## Last image crossfeeding
|
430 |
-
- branch1_crossfeed_power: Controls the level of cross-feeding between the first and last image branch. For preserving structures.
|
431 |
-
- branch1_crossfeed_range: Sets the duration of active crossfeed during development. High values enforce strong structural similarity.
|
432 |
-
- branch1_crossfeed_decay: Sets decay for branch1_crossfeed_power. Lower values make the decay stronger across the range.
|
433 |
-
|
434 |
-
## Transition settings
|
435 |
-
- parental_crossfeed_power: Similar to branch1_crossfeed_power, however applied for the images withinin the transition.
|
436 |
-
- parental_crossfeed_range: Similar to branch1_crossfeed_range, however applied for the images withinin the transition.
|
437 |
-
- parental_crossfeed_power_decay: Similar to branch1_crossfeed_decay, however applied for the images withinin the transition.
|
438 |
-
- depth_strength: Determines when the blending process will begin in terms of diffusion steps. Low values more inventive but can cause motion.
|
439 |
-
- guidance_scale_mid_damper: Decreases the guidance scale in the middle of a transition.
|
440 |
-
""")
|
441 |
-
|
442 |
-
with gr.Row():
|
443 |
-
user_id = gr.Textbox(label="user id", interactive=False)
|
444 |
-
|
445 |
-
# Collect all UI elemts in list to easily pass as inputs in gradio
|
446 |
-
dict_ui_elem = {}
|
447 |
-
dict_ui_elem["prompt1"] = prompt1
|
448 |
-
dict_ui_elem["negative_prompt"] = negative_prompt
|
449 |
-
dict_ui_elem["prompt2"] = prompt2
|
450 |
-
|
451 |
-
dict_ui_elem["duration_compute"] = duration_compute
|
452 |
-
dict_ui_elem["duration_video"] = duration_video
|
453 |
-
dict_ui_elem["height"] = height
|
454 |
-
dict_ui_elem["width"] = width
|
455 |
-
|
456 |
-
dict_ui_elem["depth_strength"] = depth_strength
|
457 |
-
dict_ui_elem["branch1_crossfeed_power"] = branch1_crossfeed_power
|
458 |
-
dict_ui_elem["branch1_crossfeed_range"] = branch1_crossfeed_range
|
459 |
-
dict_ui_elem["branch1_crossfeed_decay"] = branch1_crossfeed_decay
|
460 |
-
|
461 |
-
dict_ui_elem["num_inference_steps"] = num_inference_steps
|
462 |
-
dict_ui_elem["guidance_scale"] = guidance_scale
|
463 |
-
dict_ui_elem["guidance_scale_mid_damper"] = guidance_scale_mid_damper
|
464 |
-
dict_ui_elem["seed1"] = seed1
|
465 |
-
dict_ui_elem["seed2"] = seed2
|
466 |
-
|
467 |
-
dict_ui_elem["parental_crossfeed_range"] = parental_crossfeed_range
|
468 |
-
dict_ui_elem["parental_crossfeed_power"] = parental_crossfeed_power
|
469 |
-
dict_ui_elem["parental_crossfeed_power_decay"] = parental_crossfeed_power_decay
|
470 |
-
dict_ui_elem["user_id"] = user_id
|
471 |
-
|
472 |
-
# Convert to list, as gradio doesn't seem to accept dicts
|
473 |
-
list_ui_vals = []
|
474 |
-
list_ui_keys = []
|
475 |
-
for k in dict_ui_elem.keys():
|
476 |
-
list_ui_vals.append(dict_ui_elem[k])
|
477 |
-
list_ui_keys.append(k)
|
478 |
-
bf.list_ui_keys = list_ui_keys
|
479 |
-
|
480 |
-
b_newseed1.click(bf.randomize_seed1, outputs=seed1)
|
481 |
-
b_newseed2.click(bf.randomize_seed2, outputs=seed2)
|
482 |
-
b_compute1.click(bf.compute_img1, inputs=list_ui_vals, outputs=[img1, img2, img3, img4, img5, user_id])
|
483 |
-
b_compute2.click(bf.compute_img2, inputs=list_ui_vals, outputs=[img2, img3, img4, img5, user_id])
|
484 |
-
b_compute_transition.click(bf.compute_transition,
|
485 |
-
inputs=list_ui_vals,
|
486 |
-
outputs=[img2, img3, img4, vid_single])
|
487 |
-
|
488 |
-
b_stackforward.click(bf.stack_forward,
|
489 |
-
inputs=[prompt2, seed2],
|
490 |
-
outputs=[vid_multi, img1, img2, img3, img4, img5, prompt1, seed1, prompt2])
|
491 |
-
|
492 |
-
demo.launch(share=bf.share, inbrowser=True, inline=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|