Spaces:
Running
on
Zero
Running
on
Zero
Create app_with_diffusers.py
Browse files- app_with_diffusers.py +68 -0
app_with_diffusers.py
ADDED
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from huggingface_hub import hf_hub_download
|
2 |
+
|
3 |
+
hf_hub_download(repo_id="InstantX/InstantIR", filename="models/adapter.pt", local_dir=".")
|
4 |
+
hf_hub_download(repo_id="InstantX/InstantIR", filename="models/aggregator.pt", local_dir=".")
|
5 |
+
hf_hub_download(repo_id="InstantX/InstantIR", filename="models/previewer_lora_weights.bin", local_dir=".")
|
6 |
+
|
7 |
+
import torch
|
8 |
+
from PIL import Image
|
9 |
+
|
10 |
+
from diffusers import DDPMScheduler
|
11 |
+
from schedulers.lcm_single_step_scheduler import LCMSingleStepScheduler
|
12 |
+
|
13 |
+
from module.ip_adapter.utils import load_adapter_to_pipe
|
14 |
+
from pipelines.sdxl_instantir import InstantIRPipeline
|
15 |
+
|
16 |
+
# prepare models under ./models
|
17 |
+
instantir_path = f'./models'
|
18 |
+
|
19 |
+
# load pretrained models
|
20 |
+
pipe = InstantIRPipeline.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', torch_dtype=torch.float16)
|
21 |
+
|
22 |
+
# load adapter
|
23 |
+
load_adapter_to_pipe(
|
24 |
+
pipe,
|
25 |
+
f"{instantir_path}/adapter.pt",
|
26 |
+
image_encoder_or_path = 'facebook/dinov2-large',
|
27 |
+
)
|
28 |
+
|
29 |
+
# load previewer lora
|
30 |
+
pipe.prepare_previewers(instantir_path)
|
31 |
+
pipe.scheduler = DDPMScheduler.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', subfolder="scheduler")
|
32 |
+
lcm_scheduler = LCMSingleStepScheduler.from_config(pipe.scheduler.config)
|
33 |
+
|
34 |
+
# load aggregator weights
|
35 |
+
pretrained_state_dict = torch.load(f"{instantir_path}/aggregator.pt")
|
36 |
+
pipe.aggregator.load_state_dict(pretrained_state_dict)
|
37 |
+
|
38 |
+
# send to GPU and fp16
|
39 |
+
pipe.to(device='cuda', dtype=torch.float16)
|
40 |
+
pipe.aggregator.to(device='cuda', dtype=torch.float16)
|
41 |
+
|
42 |
+
def infer(input_image):
|
43 |
+
# load a broken image
|
44 |
+
low_quality_image = Image.open(input_image).convert("RGB")
|
45 |
+
|
46 |
+
# InstantIR restoration
|
47 |
+
image = pipe(
|
48 |
+
image=low_quality_image,
|
49 |
+
previewer_scheduler=lcm_scheduler,
|
50 |
+
).images[0]
|
51 |
+
|
52 |
+
return image
|
53 |
+
|
54 |
+
import gradio as gr
|
55 |
+
|
56 |
+
with gr.Blocks() as demo:
|
57 |
+
with gr.Column():
|
58 |
+
with gr.Row():
|
59 |
+
with gr.Column():
|
60 |
+
lq_img = gr.Image(label="Low-quality image", type="filepath")
|
61 |
+
submit_btn = gr.Button("InstantIR magic!")
|
62 |
+
output_img = gr.Image(label="InstantIR restored")
|
63 |
+
submit_btn.click(
|
64 |
+
fn=infer,
|
65 |
+
inputs=[lq_img],
|
66 |
+
outputs=[output_img]
|
67 |
+
)
|
68 |
+
demo.launch(show_error=True)
|