Linoy Tsaban
commited on
Commit
•
7b3a214
1
Parent(s):
69cb2e3
Update app.py
Browse files
app.py
CHANGED
@@ -132,7 +132,6 @@ def invert_and_reconstruct(
|
|
132 |
src_cfg_scale = 3.5,
|
133 |
skip=36,
|
134 |
tar_cfg_scale=15,
|
135 |
-
# neg_guidance=False,
|
136 |
|
137 |
):
|
138 |
|
@@ -146,16 +145,11 @@ def invert_and_reconstruct(
|
|
146 |
zs = gr.State(value=zs_tensor)
|
147 |
do_inversion = False
|
148 |
|
149 |
-
output = sample(zs.value, wts.value, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale)
|
150 |
|
151 |
-
return output, wts, zs, do_inversion
|
|
|
152 |
|
153 |
-
def update_sega_concept_table(edit_concept, neg_guidance, concepts_table):
|
154 |
-
if edit_concept:
|
155 |
-
new_rows = concepts_table.value.append([edit_concept, neg_guidance])
|
156 |
-
new_concepts_table = gr.DataFrame(value=new_rows)
|
157 |
-
return new_concepts_table
|
158 |
-
return concepts_table
|
159 |
|
160 |
def edit(input_image,
|
161 |
wts, zs,
|
@@ -163,49 +157,29 @@ def edit(input_image,
|
|
163 |
steps=100,
|
164 |
skip=36,
|
165 |
tar_cfg_scale=15,
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
|
|
170 |
|
171 |
):
|
172 |
|
173 |
# SEGA
|
174 |
# parse concepts and neg guidance
|
175 |
-
|
176 |
-
num_concepts = len(edit_concepts)
|
177 |
-
neg_guidance =[]
|
178 |
-
for edit_concept in edit_concepts:
|
179 |
-
edit_concept=edit_concept.strip(" ")
|
180 |
-
if edit_concept.startswith("-"):
|
181 |
-
neg_guidance.append(True)
|
182 |
-
else:
|
183 |
-
neg_guidance.append(False)
|
184 |
-
edit_concepts = [concept.strip("+|-") for concept in edit_concepts]
|
185 |
-
|
186 |
-
# parse warm-up steps
|
187 |
-
default_warm_up_steps = [1]*num_concepts
|
188 |
-
if warm_up:
|
189 |
-
digit_pattern = re.compile(r"^\d+$")
|
190 |
-
warm_up_steps_str = warm_up.split(",")
|
191 |
-
for i,num_steps in enumerate(warm_up_steps_str[:num_concepts]):
|
192 |
-
if not digit_pattern.match(num_steps):
|
193 |
-
raise gr.Error("Invalid value for warm-up steps, using 1 instead")
|
194 |
-
else:
|
195 |
-
default_warm_up_steps[i] = int(num_steps)
|
196 |
-
|
197 |
-
|
198 |
editing_args = dict(
|
199 |
-
editing_prompt =
|
200 |
-
reverse_editing_direction =
|
201 |
-
edit_warmup_steps=
|
202 |
-
edit_guidance_scale=[
|
203 |
-
edit_threshold=[
|
204 |
edit_momentum_scale=0.5,
|
205 |
-
edit_mom_beta=0.6
|
|
|
206 |
)
|
207 |
latnets = wts.value[skip].expand(1, -1, -1, -1)
|
208 |
-
sega_out = sem_pipe(prompt=tar_prompt,
|
209 |
num_images_per_prompt=1,
|
210 |
num_inference_steps=steps,
|
211 |
use_ddpm=True, wts=wts.value, zs=zs.value[skip:], **editing_args)
|
@@ -246,7 +220,7 @@ with gr.Blocks(css='style.css') as demo:
|
|
246 |
|
247 |
with gr.Row():
|
248 |
input_image = gr.Image(label="Input Image", interactive=True)
|
249 |
-
ddpm_edited_image = gr.Image(label=f"DDPM Reconstructed Image", interactive=False, visible=False)
|
250 |
sega_edited_image = gr.Image(label=f"DDPM + SEGA Edited Image", interactive=False)
|
251 |
input_image.style(height=512, width=512)
|
252 |
ddpm_edited_image.style(height=512, width=512)
|
@@ -254,31 +228,20 @@ with gr.Blocks(css='style.css') as demo:
|
|
254 |
|
255 |
with gr.Row():
|
256 |
tar_prompt = gr.Textbox(lines=1, label="Target Prompt", interactive=True, placeholder="")
|
257 |
-
|
258 |
-
|
259 |
-
# edit_concept = gr.Textbox(lines=1, label="Enter SEGA Edit Concept", visible = True, interactive=True)
|
260 |
-
# with gr.Column(scale=1):
|
261 |
-
# neg_guidance = gr.Checkbox(label="Negative Guidance", value=False)
|
262 |
-
# submit_concept = gr.Button(label="Add Concept")
|
263 |
-
# concepts_table = gr.Dataframe(
|
264 |
-
# headers=["Concepts", "Negative Guidance"],
|
265 |
-
# datatype=["str", "bool"],
|
266 |
-
# label="SEGA Concepts",
|
267 |
-
# )
|
268 |
-
|
269 |
-
|
270 |
with gr.Row():
|
271 |
with gr.Column(scale=1, min_width=100):
|
272 |
-
|
273 |
-
with gr.Column(scale=1, min_width=100):
|
274 |
-
|
275 |
|
276 |
with gr.Accordion("Advanced Options", open=False):
|
277 |
with gr.Tabs() as tabs:
|
278 |
with gr.TabItem('SEGA Guidance', id=0):
|
279 |
with gr.Row().style(mobile_collapse=False, equal_height=True):
|
280 |
-
|
281 |
-
label="Edit
|
282 |
show_label=False,
|
283 |
max_lines=1,
|
284 |
placeholder="Enter your 1st edit prompt",
|
@@ -289,11 +252,12 @@ with gr.Blocks(css='style.css') as demo:
|
|
289 |
)
|
290 |
with gr.Group():
|
291 |
with gr.Row().style(mobile_collapse=False, equal_height=True):
|
292 |
-
|
293 |
label='Negative Guidance')
|
294 |
warmup_1 = gr.Slider(label='Warmup', minimum=0, maximum=50, value=10, step=1, interactive=True)
|
295 |
-
|
296 |
threshold_1 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99, value=0.95, steps=0.01, interactive=True)
|
|
|
297 |
with gr.TabItem('DDPM Guidance', id=1):
|
298 |
with gr.Row():
|
299 |
with gr.Column():
|
@@ -305,17 +269,14 @@ with gr.Blocks(css='style.css') as demo:
|
|
305 |
with gr.Column():
|
306 |
skip = gr.Slider(minimum=0, maximum=40, value=36, label="Skip Steps", interactive=True)
|
307 |
tar_cfg_scale = gr.Slider(minimum=7, maximum=18,value=15, label=f"Guidance Scale", interactive=True)
|
308 |
-
sega_edit_guidance = gr.Slider(value=10, label=f"SEGA Edit Guidance Scale", interactive=True)
|
309 |
-
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...")
|
310 |
|
311 |
-
|
312 |
-
# neg_guidance = gr.Checkbox(label="SEGA Negative Guidance")
|
313 |
|
314 |
|
315 |
# gr.Markdown(help_text)
|
316 |
|
317 |
|
318 |
-
|
319 |
fn = randomize_seed_fn,
|
320 |
inputs = [seed, randomize_seed],
|
321 |
outputs = [seed],
|
@@ -332,26 +293,26 @@ with gr.Blocks(css='style.css') as demo:
|
|
332 |
skip,
|
333 |
tar_cfg_scale,
|
334 |
],
|
335 |
-
outputs=[ddpm_edited_image, wts, zs, do_inversion],
|
336 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
337 |
|
338 |
-
|
339 |
-
|
340 |
-
# inputs=[input_image,
|
341 |
-
# wts, zs,
|
342 |
-
# tar_prompt,
|
343 |
-
# steps,
|
344 |
-
# skip,
|
345 |
-
# tar_cfg_scale,
|
346 |
-
# edit_concept,
|
347 |
-
# sega_edit_guidance,
|
348 |
-
# warm_up,
|
349 |
-
# # neg_guidance,
|
350 |
-
|
351 |
-
# ],
|
352 |
-
# outputs=[sega_edited_image],
|
353 |
|
354 |
-
|
355 |
|
356 |
input_image.change(
|
357 |
fn = reset_do_inversion,
|
|
|
132 |
src_cfg_scale = 3.5,
|
133 |
skip=36,
|
134 |
tar_cfg_scale=15,
|
|
|
135 |
|
136 |
):
|
137 |
|
|
|
145 |
zs = gr.State(value=zs_tensor)
|
146 |
do_inversion = False
|
147 |
|
148 |
+
# output = sample(zs.value, wts.value, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale)
|
149 |
|
150 |
+
# return output, wts, zs, do_inversion
|
151 |
+
return wts, zs, do_inversion
|
152 |
|
|
|
|
|
|
|
|
|
|
|
|
|
153 |
|
154 |
def edit(input_image,
|
155 |
wts, zs,
|
|
|
157 |
steps=100,
|
158 |
skip=36,
|
159 |
tar_cfg_scale=15,
|
160 |
+
edit_concept_1 = "",
|
161 |
+
guidnace_scale_1 = 10,
|
162 |
+
warmup_1 = 1,
|
163 |
+
neg_guidance_1 = False,
|
164 |
+
threshold_1 = 0.95
|
165 |
|
166 |
):
|
167 |
|
168 |
# SEGA
|
169 |
# parse concepts and neg guidance
|
170 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
171 |
editing_args = dict(
|
172 |
+
editing_prompt = [edit_concept_1],
|
173 |
+
reverse_editing_direction = [neg_guidance_1],
|
174 |
+
edit_warmup_steps=[warmup_1],
|
175 |
+
edit_guidance_scale=[guidnace_scale_1],
|
176 |
+
edit_threshold=[threshold_1],
|
177 |
edit_momentum_scale=0.5,
|
178 |
+
edit_mom_beta=0.6,
|
179 |
+
eta=1,
|
180 |
)
|
181 |
latnets = wts.value[skip].expand(1, -1, -1, -1)
|
182 |
+
sega_out = sem_pipe(prompt=tar_prompt, latents=latnets, guidance_scale = tar_cfg_scale,
|
183 |
num_images_per_prompt=1,
|
184 |
num_inference_steps=steps,
|
185 |
use_ddpm=True, wts=wts.value, zs=zs.value[skip:], **editing_args)
|
|
|
220 |
|
221 |
with gr.Row():
|
222 |
input_image = gr.Image(label="Input Image", interactive=True)
|
223 |
+
# ddpm_edited_image = gr.Image(label=f"DDPM Reconstructed Image", interactive=False, visible=False)
|
224 |
sega_edited_image = gr.Image(label=f"DDPM + SEGA Edited Image", interactive=False)
|
225 |
input_image.style(height=512, width=512)
|
226 |
ddpm_edited_image.style(height=512, width=512)
|
|
|
228 |
|
229 |
with gr.Row():
|
230 |
tar_prompt = gr.Textbox(lines=1, label="Target Prompt", interactive=True, placeholder="")
|
231 |
+
|
232 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
233 |
with gr.Row():
|
234 |
with gr.Column(scale=1, min_width=100):
|
235 |
+
run_button = gr.Button("Run")
|
236 |
+
# with gr.Column(scale=1, min_width=100):
|
237 |
+
# edit_button = gr.Button("Edit")
|
238 |
|
239 |
with gr.Accordion("Advanced Options", open=False):
|
240 |
with gr.Tabs() as tabs:
|
241 |
with gr.TabItem('SEGA Guidance', id=0):
|
242 |
with gr.Row().style(mobile_collapse=False, equal_height=True):
|
243 |
+
edit_concept_1 = gr.Textbox(
|
244 |
+
label="Edit Concept",
|
245 |
show_label=False,
|
246 |
max_lines=1,
|
247 |
placeholder="Enter your 1st edit prompt",
|
|
|
252 |
)
|
253 |
with gr.Group():
|
254 |
with gr.Row().style(mobile_collapse=False, equal_height=True):
|
255 |
+
neg_guidance_1 = gr.Checkbox(
|
256 |
label='Negative Guidance')
|
257 |
warmup_1 = gr.Slider(label='Warmup', minimum=0, maximum=50, value=10, step=1, interactive=True)
|
258 |
+
guidnace_scale_1 = gr.Slider(label='Scale', minimum=1, maximum=10, value=5, step=0.25, interactive=True)
|
259 |
threshold_1 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99, value=0.95, steps=0.01, interactive=True)
|
260 |
+
|
261 |
with gr.TabItem('DDPM Guidance', id=1):
|
262 |
with gr.Row():
|
263 |
with gr.Column():
|
|
|
269 |
with gr.Column():
|
270 |
skip = gr.Slider(minimum=0, maximum=40, value=36, label="Skip Steps", interactive=True)
|
271 |
tar_cfg_scale = gr.Slider(minimum=7, maximum=18,value=15, label=f"Guidance Scale", interactive=True)
|
|
|
|
|
272 |
|
273 |
+
|
|
|
274 |
|
275 |
|
276 |
# gr.Markdown(help_text)
|
277 |
|
278 |
|
279 |
+
run_button.click(
|
280 |
fn = randomize_seed_fn,
|
281 |
inputs = [seed, randomize_seed],
|
282 |
outputs = [seed],
|
|
|
293 |
skip,
|
294 |
tar_cfg_scale,
|
295 |
],
|
296 |
+
# outputs=[ddpm_edited_image, wts, zs, do_inversion],
|
297 |
+
outputs=[wts, zs, do_inversion],
|
298 |
+
).success(
|
299 |
+
fn=edit,
|
300 |
+
inputs=[input_image,
|
301 |
+
wts, zs,
|
302 |
+
tar_prompt,
|
303 |
+
steps,
|
304 |
+
skip,
|
305 |
+
tar_cfg_scale,
|
306 |
+
edit_concept_1,
|
307 |
+
guidnace_scale_1,
|
308 |
+
warmup_1,
|
309 |
+
neg_guidance_1,
|
310 |
+
threshold_1
|
311 |
|
312 |
+
],
|
313 |
+
outputs=[sega_edited_image],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
314 |
|
315 |
+
)
|
316 |
|
317 |
input_image.change(
|
318 |
fn = reset_do_inversion,
|