Spaces:
Paused
Paused
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import os
|
3 |
+
from diffusers import DiffusionPipeline, FluxControlPipeline, FluxTransformer2DModel
|
4 |
+
import torch
|
5 |
+
from transformers import T5EncoderModel
|
6 |
+
from controlnet_aux import CannyDetector
|
7 |
+
from diffusers.utils import load_image
|
8 |
+
|
9 |
+
from huggingface_hub import login
|
10 |
+
hf_token = os.environ.get["HF_TOKEN"]
|
11 |
+
login(hf_token)
|
12 |
+
|
13 |
+
def load_pipeline(four_bit=False):
|
14 |
+
orig_pipeline = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)
|
15 |
+
if four_bit:
|
16 |
+
print("Using four bit.")
|
17 |
+
transformer = FluxTransformer2DModel.from_pretrained(
|
18 |
+
"sayakpaul/FLUX.1-Canny-dev-nf4", subfolder="transformer", torch_dtype=torch.bfloat16
|
19 |
+
)
|
20 |
+
text_encoder_2 = T5EncoderModel.from_pretrained(
|
21 |
+
"sayakpaul/FLUX.1-Canny-dev-nf4", subfolder="text_encoder_2", torch_dtype=torch.bfloat16
|
22 |
+
)
|
23 |
+
pipeline = FluxControlPipeline.from_pipe(
|
24 |
+
orig_pipeline, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16
|
25 |
+
)
|
26 |
+
else:
|
27 |
+
transformer = FluxTransformer2DModel.from_pretrained(
|
28 |
+
"black-forest-labs/FLUX.1-Canny-dev",
|
29 |
+
subfolder="transformer",
|
30 |
+
revision="refs/pr/1",
|
31 |
+
torch_dtype=torch.bfloat16,
|
32 |
+
)
|
33 |
+
pipeline = FluxControlPipeline.from_pipe(orig_pipeline, transformer=transformer, torch_dtype=torch.bfloat16)
|
34 |
+
|
35 |
+
pipeline.enable_model_cpu_offload()
|
36 |
+
return pipeline
|
37 |
+
|
38 |
+
def get_canny(control_image):
|
39 |
+
processor = CannyDetector()
|
40 |
+
control_image = processor(
|
41 |
+
control_image, low_threshold=50, high_threshold=200, detect_resolution=1024, image_resolution=1024
|
42 |
+
)
|
43 |
+
return control_image
|
44 |
+
|
45 |
+
def main(ref_filepath, prompt, use_nf4):
|
46 |
+
pipe = load_pipeline(use_nf4)
|
47 |
+
control_image = load_image(ref_filepath)
|
48 |
+
control_image = get_canny(control_image)
|
49 |
+
image = pipe(
|
50 |
+
prompt=prompt,
|
51 |
+
control_image=control_image,
|
52 |
+
height=1024,
|
53 |
+
width=1024,
|
54 |
+
num_inference_steps=50,
|
55 |
+
guidance_scale=30.0,
|
56 |
+
max_sequence_length=512,
|
57 |
+
generator=torch.Generator("cpu").manual_seed(0),
|
58 |
+
).images[0]
|
59 |
+
filename = "output_"
|
60 |
+
filename += "_4bit" if four_bit else ""
|
61 |
+
image.save(f"{filename}.png")
|
62 |
+
return f"{filename}.png", control_image
|
63 |
+
|
64 |
+
with gr.Blocks() as demo:
|
65 |
+
with gr.Column():
|
66 |
+
gr.Markdown("# FLUX.1 Canny Dev")
|
67 |
+
with gr.Row():
|
68 |
+
with gr.Column():
|
69 |
+
image_input = gr.Image(label="Reference Image", type="filepath")
|
70 |
+
prompt = gr.Textbox(label="Prompt")
|
71 |
+
use_nf4 = gr.Checkbox(label="Use NF4 checkpoints", value=True)
|
72 |
+
submit_btn = gr.Button("Submit")
|
73 |
+
with gr.Column():
|
74 |
+
results= gr.Gallery(label="Results")
|
75 |
+
|
76 |
+
submit_btn.click(
|
77 |
+
fn = main,
|
78 |
+
inputs = [image_input, prompt, use_nf4],
|
79 |
+
outputs = [results]
|
80 |
+
)
|
81 |
+
|
82 |
+
demo.launch(show_api=False, show_error=True)
|