linoyts HF staff commited on
Commit
592c68b
Β·
verified Β·
1 Parent(s): 292c38f

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +96 -3
app.py CHANGED
@@ -25,10 +25,16 @@ pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell",
25
 
26
  pipe.transformer.to(memory_format=torch.channels_last)
27
  pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
28
-
29
  clip_slider = CLIPSliderFlux(pipe, device=torch.device("cuda"))
30
 
31
 
 
 
 
 
 
 
32
  @spaces.GPU(duration=200)
33
  def generate(slider_x, prompt, seed, recalc_directions, iterations, steps, guidance_scale,
34
  x_concept_1, x_concept_2,
@@ -41,6 +47,7 @@ def generate(slider_x, prompt, seed, recalc_directions, iterations, steps, guida
41
  # check if avg diff for directions need to be re-calculated
42
  print("slider_x", slider_x)
43
  print("x_concept_1", x_concept_1, "x_concept_2", x_concept_2)
 
44
 
45
  if not sorted(slider_x) == sorted([x_concept_1, x_concept_2]) or recalc_directions:
46
  #avg_diff = clip_slider.find_latent_direction(slider_x[0], slider_x[1], num_iterations=iterations).to(torch.float16)
@@ -59,6 +66,7 @@ def generate(slider_x, prompt, seed, recalc_directions, iterations, steps, guida
59
  seed=seed, num_inference_steps=steps, avg_diff=avg_diff)
60
 
61
 
 
62
  comma_concepts_x = f"{slider_x[1]}, {slider_x[0]}"
63
 
64
  avg_diff_x = avg_diff.cpu()
@@ -71,6 +79,7 @@ def update_scales(x,prompt,seed, steps, guidance_scale,
71
  img2img_type = None, img = None,
72
  controlnet_scale= None, ip_adapter_scale=None,):
73
  avg_diff = avg_diff_x.cuda()
 
74
  if img2img_type=="controlnet canny" and img is not None:
75
  control_img = process_controlnet_img(img)
76
  image = t5_slider_controlnet.generate(prompt, guidance_scale=guidance_scale, image=control_img, controlnet_conditioning_scale =controlnet_scale, scale=x, seed=seed, num_inference_steps=steps, avg_diff=avg_diff)
@@ -81,6 +90,27 @@ def update_scales(x,prompt,seed, steps, guidance_scale,
81
  return image
82
 
83
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
84
  def reset_recalc_directions():
85
  return True
86
 
@@ -131,7 +161,10 @@ intro = """
131
  </p>
132
  """
133
  with gr.Blocks(css=css) as demo:
134
-
 
 
 
135
  gr.HTML(intro)
136
 
137
  x_concept_1 = gr.State("")
@@ -144,6 +177,7 @@ with gr.Blocks(css=css) as demo:
144
 
145
  recalc_directions = gr.State(False)
146
 
 
147
  with gr.Row():
148
  with gr.Column():
149
  slider_x = gr.Dropdown(label="Slider concept range", allow_custom_value=True, multiselect=True, max_choices=2)
@@ -168,9 +202,63 @@ with gr.Blocks(css=css) as demo:
168
  step=0.1,
169
  value=5,
170
  )
171
-
 
 
 
 
 
 
172
  seed = gr.Slider(minimum=0, maximum=np.iinfo(np.int32).max, label="Seed", interactive=True, randomize=True)
173
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
174
  submit.click(fn=generate,
175
  inputs=[slider_x, prompt, seed, recalc_directions, iterations, steps, guidance_scale, x_concept_1, x_concept_2, avg_diff_x],
176
  outputs=[x, x_concept_1, x_concept_2, avg_diff_x, output_image])
@@ -178,6 +266,11 @@ with gr.Blocks(css=css) as demo:
178
  iterations.change(fn=reset_recalc_directions, outputs=[recalc_directions])
179
  seed.change(fn=reset_recalc_directions, outputs=[recalc_directions])
180
  x.change(fn=update_scales, inputs=[x, prompt, seed, steps, guidance_scale, avg_diff_x], outputs=[output_image])
 
 
 
 
 
181
 
182
  if __name__ == "__main__":
183
  demo.launch()
 
25
 
26
  pipe.transformer.to(memory_format=torch.channels_last)
27
  pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
28
+ #pipe.enable_model_cpu_offload()
29
  clip_slider = CLIPSliderFlux(pipe, device=torch.device("cuda"))
30
 
31
 
32
+ base_model = 'black-forest-labs/FLUX.1-schnell'
33
+ controlnet_model = 'InstantX/FLUX.1-dev-Controlnet-Canny-alpha'
34
+ # controlnet = FluxControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.bfloat16)
35
+ # pipe_controlnet = FluxControlNetPipeline.from_pretrained(base_model, controlnet=controlnet, torch_dtype=torch.bfloat16)
36
+ # t5_slider_controlnet = T5SliderFlux(sd_pipe=pipe_controlnet,device=torch.device("cuda"))
37
+
38
  @spaces.GPU(duration=200)
39
  def generate(slider_x, prompt, seed, recalc_directions, iterations, steps, guidance_scale,
40
  x_concept_1, x_concept_2,
 
47
  # check if avg diff for directions need to be re-calculated
48
  print("slider_x", slider_x)
49
  print("x_concept_1", x_concept_1, "x_concept_2", x_concept_2)
50
+ #torch.manual_seed(seed)
51
 
52
  if not sorted(slider_x) == sorted([x_concept_1, x_concept_2]) or recalc_directions:
53
  #avg_diff = clip_slider.find_latent_direction(slider_x[0], slider_x[1], num_iterations=iterations).to(torch.float16)
 
66
  seed=seed, num_inference_steps=steps, avg_diff=avg_diff)
67
 
68
 
69
+ #comma_concepts_x = ', '.join(slider_x)
70
  comma_concepts_x = f"{slider_x[1]}, {slider_x[0]}"
71
 
72
  avg_diff_x = avg_diff.cpu()
 
79
  img2img_type = None, img = None,
80
  controlnet_scale= None, ip_adapter_scale=None,):
81
  avg_diff = avg_diff_x.cuda()
82
+ torch.manual_seed(seed)
83
  if img2img_type=="controlnet canny" and img is not None:
84
  control_img = process_controlnet_img(img)
85
  image = t5_slider_controlnet.generate(prompt, guidance_scale=guidance_scale, image=control_img, controlnet_conditioning_scale =controlnet_scale, scale=x, seed=seed, num_inference_steps=steps, avg_diff=avg_diff)
 
90
  return image
91
 
92
 
93
+
94
+ @spaces.GPU
95
+ def update_x(x,y,prompt,seed, steps,
96
+ avg_diff_x, avg_diff_y,
97
+ img2img_type = None,
98
+ img = None):
99
+ avg_diff = avg_diff_x.cuda()
100
+ avg_diff_2nd = avg_diff_y.cuda()
101
+ image = clip_slider.generate(prompt, scale=x, scale_2nd=y, seed=seed, num_inference_steps=steps, avg_diff=avg_diff,avg_diff_2nd=avg_diff_2nd)
102
+ return image
103
+
104
+ @spaces.GPU
105
+ def update_y(x,y,prompt,seed, steps,
106
+ avg_diff_x, avg_diff_y,
107
+ img2img_type = None,
108
+ img = None):
109
+ avg_diff = avg_diff_x.cuda()
110
+ avg_diff_2nd = avg_diff_y.cuda()
111
+ image = clip_slider.generate(prompt, scale=x, scale_2nd=y, seed=seed, num_inference_steps=steps, avg_diff=avg_diff,avg_diff_2nd=avg_diff_2nd)
112
+ return image
113
+
114
  def reset_recalc_directions():
115
  return True
116
 
 
161
  </p>
162
  """
163
  with gr.Blocks(css=css) as demo:
164
+ # gr.Markdown(f"""# Latent Navigation
165
+ # ## Exploring CLIP text space with FLUX.1 schnell πŸͺ
166
+ # [[code](https://github.com/linoytsaban/semantic-sliders)]
167
+ # """)
168
  gr.HTML(intro)
169
 
170
  x_concept_1 = gr.State("")
 
177
 
178
  recalc_directions = gr.State(False)
179
 
180
+ #with gr.Tab("text2image"):
181
  with gr.Row():
182
  with gr.Column():
183
  slider_x = gr.Dropdown(label="Slider concept range", allow_custom_value=True, multiselect=True, max_choices=2)
 
202
  step=0.1,
203
  value=5,
204
  )
205
+ # correlation = gr.Slider(
206
+ # label="correlation",
207
+ # minimum=0.1,
208
+ # maximum=1.0,
209
+ # step=0.05,
210
+ # value=0.6,
211
+ # )
212
  seed = gr.Slider(minimum=0, maximum=np.iinfo(np.int32).max, label="Seed", interactive=True, randomize=True)
213
 
214
+
215
+ # with gr.Tab(label="image2image"):
216
+ # with gr.Row():
217
+ # with gr.Column():
218
+ # image = gr.ImageEditor(type="pil", image_mode="L", crop_size=(512, 512))
219
+ # slider_x_a = gr.Dropdown(label="Slider X concept range", allow_custom_value=True, multiselect=True, max_choices=2)
220
+ # slider_y_a = gr.Dropdown(label="Slider X concept range", allow_custom_value=True, multiselect=True, max_choices=2)
221
+ # img2img_type = gr.Radio(["controlnet canny", "ip adapter"], label="", info="", visible=False, value="controlnet canny")
222
+ # prompt_a = gr.Textbox(label="Prompt")
223
+ # submit_a = gr.Button("Submit")
224
+ # with gr.Column():
225
+ # with gr.Group(elem_id="group"):
226
+ # x_a = gr.Slider(minimum=-10, value=0, maximum=10, elem_id="x", interactive=False)
227
+ # y_a = gr.Slider(minimum=-10, value=0, maximum=10, elem_id="y", interactive=False)
228
+ # output_image_a = gr.Image(elem_id="image_out")
229
+ # with gr.Row():
230
+ # generate_butt_a = gr.Button("generate")
231
+
232
+ # with gr.Accordion(label="advanced options", open=False):
233
+ # iterations_a = gr.Slider(label = "num iterations", minimum=0, value=200, maximum=300)
234
+ # steps_a = gr.Slider(label = "num inference steps", minimum=1, value=8, maximum=30)
235
+ # guidance_scale_a = gr.Slider(
236
+ # label="Guidance scale",
237
+ # minimum=0.1,
238
+ # maximum=10.0,
239
+ # step=0.1,
240
+ # value=5,
241
+ # )
242
+ # controlnet_conditioning_scale = gr.Slider(
243
+ # label="controlnet conditioning scale",
244
+ # minimum=0.5,
245
+ # maximum=5.0,
246
+ # step=0.1,
247
+ # value=0.7,
248
+ # )
249
+ # ip_adapter_scale = gr.Slider(
250
+ # label="ip adapter scale",
251
+ # minimum=0.5,
252
+ # maximum=5.0,
253
+ # step=0.1,
254
+ # value=0.8,
255
+ # visible=False
256
+ # )
257
+ # seed_a = gr.Slider(minimum=0, maximum=np.iinfo(np.int32).max, label="Seed", interactive=True, randomize=True)
258
+
259
+ # submit.click(fn=generate,
260
+ # inputs=[slider_x, slider_y, prompt, seed, iterations, steps, guidance_scale, x_concept_1, x_concept_2, y_concept_1, y_concept_2, avg_diff_x, avg_diff_y],
261
+ # outputs=[x, y, x_concept_1, x_concept_2, y_concept_1, y_concept_2, avg_diff_x, avg_diff_y, output_image])
262
  submit.click(fn=generate,
263
  inputs=[slider_x, prompt, seed, recalc_directions, iterations, steps, guidance_scale, x_concept_1, x_concept_2, avg_diff_x],
264
  outputs=[x, x_concept_1, x_concept_2, avg_diff_x, output_image])
 
266
  iterations.change(fn=reset_recalc_directions, outputs=[recalc_directions])
267
  seed.change(fn=reset_recalc_directions, outputs=[recalc_directions])
268
  x.change(fn=update_scales, inputs=[x, prompt, seed, steps, guidance_scale, avg_diff_x], outputs=[output_image])
269
+ # generate_butt_a.click(fn=update_scales, inputs=[x_a,y_a, prompt_a, seed_a, steps_a, guidance_scale_a, avg_diff_x, avg_diff_y, img2img_type, image, controlnet_conditioning_scale, ip_adapter_scale], outputs=[output_image_a])
270
+ # submit_a.click(fn=generate,
271
+ # inputs=[slider_x_a, slider_y_a, prompt_a, seed_a, iterations_a, steps_a, guidance_scale_a, x_concept_1, x_concept_2, y_concept_1, y_concept_2, avg_diff_x, avg_diff_y, img2img_type, image, controlnet_conditioning_scale, ip_adapter_scale],
272
+ # outputs=[x_a, y_a, x_concept_1, x_concept_2, y_concept_1, y_concept_2, avg_diff_x, avg_diff_y, output_image_a])
273
+
274
 
275
  if __name__ == "__main__":
276
  demo.launch()