renyuxi commited on
Commit
bad1b06
1 Parent(s): a94f1ca

Create app.py

Browse files
Files changed (1) hide show
  1. app.py +77 -0
app.py ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import os
3
+ import time
4
+ from os import path
5
+ from safetensors.torch import load_file
6
+
7
+ cache_path = path.join(path.dirname(path.abspath(__file__)), "models")
8
+ os.environ["TRANSFORMERS_CACHE"] = cache_path
9
+ os.environ["HF_HUB_CACHE"] = cache_path
10
+ os.environ["HF_HOME"] = cache_path
11
+
12
+ import gradio as gr
13
+ import torch
14
+ from diffusers import StableDiffusionXLPipeline, LCMScheduler
15
+
16
+ # from scheduling_tcd import TCDScheduler
17
+
18
+ torch.backends.cuda.matmul.allow_tf32 = True
19
+
20
+ class timer:
21
+ def __init__(self, method_name="timed process"):
22
+ self.method = method_name
23
+
24
+ def __enter__(self):
25
+ self.start = time.time()
26
+ print(f"{self.method} starts")
27
+
28
+ def __exit__(self, exc_type, exc_val, exc_tb):
29
+ end = time.time()
30
+ print(f"{self.method} took {str(round(end - self.start, 2))}s")
31
+
32
+ if not path.exists(cache_path):
33
+ os.makedirs(cache_path, exist_ok=True)
34
+
35
+ pipe = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.bfloat16, safety_checker=None)
36
+ pipe.to(device="cuda", dtype=torch.bfloat16)
37
+ unet_state = load_file(hf_hub_download("ByteDance/Hyper-SD", "Hyper-SDXL-1step-Unet.safetensors"), device="cuda")
38
+ pipe.unet.load_state_dict(unet_state)
39
+ pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config, timestep_spacing ="trailing")
40
+
41
+ with gr.Blocks() as demo:
42
+ with gr.Column():
43
+ with gr.Row():
44
+ with gr.Column():
45
+ num_images = gr.Slider(label="Number of Images", minimum=1, maximum=8, step=1, value=4, interactive=True)
46
+ height = gr.Number(label="Image Height", value=1024, interactive=True)
47
+ width = gr.Number(label="Image Width", value=1024, interactive=True)
48
+ # steps = gr.Slider(label="Inference Steps", minimum=1, maximum=8, step=1, value=1, interactive=True)
49
+ # eta = gr.Number(label="Eta (Corresponds to parameter eta (η) in the DDIM paper, i.e. 0.0 eqauls DDIM, 1.0 equals LCM)", value=1., interactive=True)
50
+ prompt = gr.Text(label="Prompt", value="a photo of a cat", interactive=True)
51
+ seed = gr.Number(label="Seed", value=3413, interactive=True)
52
+ btn = gr.Button(value="run")
53
+ with gr.Column():
54
+ output = gr.Gallery(width=1024, height=768)
55
+
56
+ def process_image(num_images, height, width, prompt, seed):
57
+ global pipe
58
+ with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16), timer("inference"):
59
+ return pipe(
60
+ prompt=[prompt]*num_images,
61
+ generator=torch.Generator().manual_seed(int(seed)),
62
+ num_inference_steps=1,
63
+ guidance_scale=0.,
64
+ height=int(height),
65
+ width=int(width),
66
+ timesteps=[800]
67
+ ).images
68
+
69
+ reactive_controls = [num_images, height, width, prompt, seed]
70
+
71
+ # for control in reactive_controls:
72
+ # control.change(fn=process_image, inputs=reactive_controls, outputs=[output])
73
+
74
+ btn.click(process_image, inputs=reactive_controls, outputs=[output])
75
+
76
+ if __name__ == "__main__":
77
+ demo.launch()