Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -25,6 +25,7 @@ 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 |
clip_slider = CLIPSliderFlux(pipe, device=torch.device("cuda"))
|
| 29 |
|
| 30 |
|
|
@@ -46,6 +47,7 @@ def generate(slider_x, prompt, seed, recalc_directions, iterations, steps, guida
|
|
| 46 |
# check if avg diff for directions need to be re-calculated
|
| 47 |
print("slider_x", slider_x)
|
| 48 |
print("x_concept_1", x_concept_1, "x_concept_2", x_concept_2)
|
|
|
|
| 49 |
|
| 50 |
if not sorted(slider_x) == sorted([x_concept_1, x_concept_2]) or recalc_directions:
|
| 51 |
#avg_diff = clip_slider.find_latent_direction(slider_x[0], slider_x[1], num_iterations=iterations).to(torch.float16)
|
|
@@ -63,6 +65,8 @@ def generate(slider_x, prompt, seed, recalc_directions, iterations, steps, guida
|
|
| 63 |
scale=0, scale_2nd=0,
|
| 64 |
seed=seed, num_inference_steps=steps, avg_diff=avg_diff)
|
| 65 |
|
|
|
|
|
|
|
| 66 |
comma_concepts_x = f"{slider_x[1]}, {slider_x[0]}"
|
| 67 |
|
| 68 |
avg_diff_x = avg_diff.cpu()
|
|
@@ -75,16 +79,36 @@ def update_scales(x,prompt,seed, steps, guidance_scale,
|
|
| 75 |
img2img_type = None, img = None,
|
| 76 |
controlnet_scale= None, ip_adapter_scale=None,):
|
| 77 |
avg_diff = avg_diff_x.cuda()
|
|
|
|
| 78 |
if img2img_type=="controlnet canny" and img is not None:
|
| 79 |
control_img = process_controlnet_img(img)
|
| 80 |
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 |
elif img2img_type=="ip adapter" and img is not None:
|
| 82 |
image = clip_slider.generate(prompt, guidance_scale=guidance_scale, ip_adapter_image=img, scale=x,seed=seed, num_inference_steps=steps, avg_diff=avg_diff)
|
| 83 |
else:
|
| 84 |
-
image = clip_slider.generate(prompt,
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
return image
|
| 89 |
|
| 90 |
def reset_recalc_directions():
|
|
@@ -98,14 +122,12 @@ css = '''
|
|
| 98 |
margin-bottom: 20px;
|
| 99 |
background-color: white;
|
| 100 |
}
|
| 101 |
-
|
| 102 |
#x {
|
| 103 |
position: absolute;
|
| 104 |
bottom: 20px; /* Moved further down */
|
| 105 |
left: 30px; /* Adjusted left margin */
|
| 106 |
width: 540px; /* Increased width to match the new container size */
|
| 107 |
}
|
| 108 |
-
|
| 109 |
#y {
|
| 110 |
position: absolute;
|
| 111 |
bottom: 200px; /* Increased bottom margin to ensure proper spacing from #x */
|
|
@@ -114,14 +136,12 @@ css = '''
|
|
| 114 |
transform: rotate(-90deg);
|
| 115 |
transform-origin: left bottom;
|
| 116 |
}
|
| 117 |
-
|
| 118 |
#image_out {
|
| 119 |
position: absolute;
|
| 120 |
width: 80%; /* Adjust width as needed */
|
| 121 |
right: 10px;
|
| 122 |
top: 10px; /* Increased top margin to clear space occupied by #x */
|
| 123 |
}
|
| 124 |
-
|
| 125 |
'''
|
| 126 |
intro = """
|
| 127 |
<div style="display: flex;align-items: center;justify-content: center">
|
|
@@ -166,7 +186,7 @@ with gr.Blocks(css=css) as demo:
|
|
| 166 |
submit = gr.Button("find directions")
|
| 167 |
with gr.Column():
|
| 168 |
with gr.Group(elem_id="group"):
|
| 169 |
-
x = gr.Slider(minimum=-3, value=0, maximum=3.5,
|
| 170 |
#y = gr.Slider(minimum=-10, value=0, maximum=10, elem_id="y", interactive=False)
|
| 171 |
output_image = gr.Image(elem_id="image_out")
|
| 172 |
# with gr.Row():
|
|
|
|
| 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 |
|
|
|
|
| 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)
|
|
|
|
| 65 |
scale=0, scale_2nd=0,
|
| 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)
|
| 86 |
elif img2img_type=="ip adapter" and img is not None:
|
| 87 |
image = clip_slider.generate(prompt, guidance_scale=guidance_scale, ip_adapter_image=img, scale=x,seed=seed, num_inference_steps=steps, avg_diff=avg_diff)
|
| 88 |
else:
|
| 89 |
+
image = clip_slider.generate(prompt, guidance_scale=guidance_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():
|
|
|
|
| 122 |
margin-bottom: 20px;
|
| 123 |
background-color: white;
|
| 124 |
}
|
|
|
|
| 125 |
#x {
|
| 126 |
position: absolute;
|
| 127 |
bottom: 20px; /* Moved further down */
|
| 128 |
left: 30px; /* Adjusted left margin */
|
| 129 |
width: 540px; /* Increased width to match the new container size */
|
| 130 |
}
|
|
|
|
| 131 |
#y {
|
| 132 |
position: absolute;
|
| 133 |
bottom: 200px; /* Increased bottom margin to ensure proper spacing from #x */
|
|
|
|
| 136 |
transform: rotate(-90deg);
|
| 137 |
transform-origin: left bottom;
|
| 138 |
}
|
|
|
|
| 139 |
#image_out {
|
| 140 |
position: absolute;
|
| 141 |
width: 80%; /* Adjust width as needed */
|
| 142 |
right: 10px;
|
| 143 |
top: 10px; /* Increased top margin to clear space occupied by #x */
|
| 144 |
}
|
|
|
|
| 145 |
'''
|
| 146 |
intro = """
|
| 147 |
<div style="display: flex;align-items: center;justify-content: center">
|
|
|
|
| 186 |
submit = gr.Button("find directions")
|
| 187 |
with gr.Column():
|
| 188 |
with gr.Group(elem_id="group"):
|
| 189 |
+
x = gr.Slider(minimum=-3, value=0, maximum=3.5, elem_id="x", interactive=False)
|
| 190 |
#y = gr.Slider(minimum=-10, value=0, maximum=10, elem_id="y", interactive=False)
|
| 191 |
output_image = gr.Image(elem_id="image_out")
|
| 192 |
# with gr.Row():
|