Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,7 +1,7 @@
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
import spaces
|
4 |
-
from diffusers import StableDiffusionXLPipeline, EulerDiscreteScheduler
|
5 |
from huggingface_hub import hf_hub_download
|
6 |
from safetensors.torch import load_file
|
7 |
|
@@ -9,26 +9,25 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
9 |
base = "stabilityai/stable-diffusion-xl-base-1.0"
|
10 |
repo = "ByteDance/SDXL-Lightning"
|
11 |
opts = {
|
12 |
-
"1 Step" :
|
13 |
-
"2 Steps" :
|
14 |
-
"4 Steps" :
|
15 |
-
"8 Steps" :
|
16 |
}
|
17 |
|
|
|
|
|
|
|
18 |
pipe = StableDiffusionXLPipeline.from_pretrained(base, torch_dtype=torch.float16, variant="fp16").to(device)
|
19 |
-
last_step = None
|
20 |
|
21 |
-
# Function
|
22 |
@spaces.GPU(enable_queue=True)
|
23 |
def generate_image(prompt, option):
|
24 |
ckpt, step = opts[option]
|
25 |
-
if
|
26 |
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", prediction_type="sample" if step == 1 else "epsilon")
|
27 |
pipe.unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device=device))
|
28 |
-
|
29 |
-
|
30 |
-
return image
|
31 |
-
|
32 |
|
33 |
with gr.Blocks() as demo:
|
34 |
gr.HTML("<h1><center>SDXL-Lightning</center></h1>")
|
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
import spaces
|
4 |
+
from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel, EulerDiscreteScheduler
|
5 |
from huggingface_hub import hf_hub_download
|
6 |
from safetensors.torch import load_file
|
7 |
|
|
|
9 |
base = "stabilityai/stable-diffusion-xl-base-1.0"
|
10 |
repo = "ByteDance/SDXL-Lightning"
|
11 |
opts = {
|
12 |
+
"1 Step" : ("sdxl_lightning_1step_unet_x0.safetensors", 1),
|
13 |
+
"2 Steps" : ("sdxl_lightning_2step_unet.safetensors", 2),
|
14 |
+
"4 Steps" : ("sdxl_lightning_4step_unet.safetensors", 4),
|
15 |
+
"8 Steps" : ("sdxl_lightning_8step_unet.safetensors", 8),
|
16 |
}
|
17 |
|
18 |
+
step_loaded = 4
|
19 |
+
unet = UNet2DConditionModel.from_config(base, subfolder="unet").to("cuda", torch.float16)
|
20 |
+
unet.load_state_dict(load_file(hf_hub_download(repo, opts["4 Steps"][0]), device=device))
|
21 |
pipe = StableDiffusionXLPipeline.from_pretrained(base, torch_dtype=torch.float16, variant="fp16").to(device)
|
|
|
22 |
|
|
|
23 |
@spaces.GPU(enable_queue=True)
|
24 |
def generate_image(prompt, option):
|
25 |
ckpt, step = opts[option]
|
26 |
+
if step_loaded != step:
|
27 |
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", prediction_type="sample" if step == 1 else "epsilon")
|
28 |
pipe.unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device=device))
|
29 |
+
step_loaded = step
|
30 |
+
return pipe(prompt, num_inference_steps=step, guidance_scale=0).images[0]
|
|
|
|
|
31 |
|
32 |
with gr.Blocks() as demo:
|
33 |
gr.HTML("<h1><center>SDXL-Lightning</center></h1>")
|