Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,76 +1,120 @@
|
|
1 |
-
import
|
2 |
import torch
|
|
|
3 |
from diffusers import DiffusionPipeline
|
4 |
-
|
|
|
|
|
|
|
|
|
5 |
|
6 |
-
#
|
|
|
7 |
def load_pipeline():
|
|
|
8 |
base_model = "black-forest-labs/FLUX.1-dev"
|
9 |
-
lora_repo = "strangerzonehf/Flux-Super-Realism-LoRA"
|
10 |
-
trigger_word = "Super Realism" # Recommended trigger word
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
pipe = DiffusionPipeline.from_pretrained(
|
15 |
base_model,
|
16 |
-
torch_dtype=torch.bfloat16
|
17 |
-
|
18 |
)
|
19 |
-
|
20 |
-
# Load
|
|
|
21 |
pipe.load_lora_weights(lora_repo)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
|
23 |
-
|
24 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
25 |
-
pipe.to(device)
|
26 |
-
return pipe
|
27 |
-
|
28 |
-
# Instantiate the pipeline once on Space startup
|
29 |
-
pipe = load_pipeline()
|
30 |
-
|
31 |
-
# Define a function for image generation
|
32 |
-
def generate_image(prompt, seed, width, height, guidance_scale, randomize_seed):
|
33 |
-
# If randomize_seed is selected, allow the model to generate a random seed
|
34 |
-
if randomize_seed:
|
35 |
-
seed = None
|
36 |
|
37 |
-
|
38 |
-
|
39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
)
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
iface = gr.Interface(
|
53 |
-
fn=generate_image,
|
54 |
-
inputs=[
|
55 |
-
gr.Textbox(
|
56 |
-
lines=2,
|
57 |
-
label="Prompt",
|
58 |
-
placeholder="Enter your prompt, e.g., 'A tiny astronaut hatching from an egg on the moon, 4k, planet theme'"
|
59 |
-
),
|
60 |
-
gr.Slider(0, 10000, step=1, value=0, label="Seed (0 for random)"),
|
61 |
-
gr.Slider(256, 1024, step=64, value=1024, label="Width"),
|
62 |
-
gr.Slider(256, 1024, step=64, value=1024, label="Height"),
|
63 |
-
gr.Slider(1, 20, step=0.5, value=6, label="Guidance Scale"),
|
64 |
-
gr.Checkbox(value=True, label="Randomize Seed")
|
65 |
-
],
|
66 |
-
outputs=gr.Image(type="pil", label="Generated Image"),
|
67 |
-
title="Flux Super Realism LoRA Demo",
|
68 |
-
description=(
|
69 |
-
"This demo uses the Flux Super Realism LoRA model for ultra-realistic image generation. "
|
70 |
-
"You can use the trigger word 'Super Realism' (recommended) along with other realism-related words "
|
71 |
-
"to guide the generation process."
|
72 |
-
),
|
73 |
-
)
|
74 |
|
75 |
-
|
76 |
-
|
|
|
|
1 |
+
import gradio as gr
|
2 |
import torch
|
3 |
+
import numpy as np
|
4 |
from diffusers import DiffusionPipeline
|
5 |
+
from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
|
6 |
+
from functools import lru_cache
|
7 |
+
from PIL import Image
|
8 |
+
from torchvision import transforms
|
9 |
+
|
10 |
|
11 |
+
# Cache pipeline loading to improve performance
|
12 |
+
@lru_cache(maxsize=1)
|
13 |
def load_pipeline():
|
14 |
+
# Load base model
|
15 |
base_model = "black-forest-labs/FLUX.1-dev"
|
|
|
|
|
|
|
|
|
|
|
16 |
pipe = DiffusionPipeline.from_pretrained(
|
17 |
base_model,
|
18 |
+
torch_dtype=torch.bfloat16
|
|
|
19 |
)
|
20 |
+
|
21 |
+
# Load LoRA weights
|
22 |
+
lora_repo = "strangerzonehf/Flux-Super-Realism-LoRA"
|
23 |
pipe.load_lora_weights(lora_repo)
|
24 |
+
|
25 |
+
# Load safety checker
|
26 |
+
safety_checker = StableDiffusionSafetyChecker.from_pretrained(
|
27 |
+
"CompVis/stable-diffusion-safety-checker"
|
28 |
+
)
|
29 |
+
feature_extractor = CLIPFeatureExtractor.from_pretrained(
|
30 |
+
"openai/clip-vit-base-patch32"
|
31 |
+
)
|
32 |
+
|
33 |
+
# Optimizations
|
34 |
+
pipe.enable_xformers_memory_efficient_attention()
|
35 |
+
pipe = pipe.to("cuda")
|
36 |
+
|
37 |
+
return pipe, safety_checker, feature_extractor
|
38 |
|
39 |
+
pipe, safety_checker, feature_extractor = load_pipeline()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
|
41 |
+
def generate_image(
|
42 |
+
prompt,
|
43 |
+
seed=42,
|
44 |
+
width=1024,
|
45 |
+
height=1024,
|
46 |
+
guidance_scale=6,
|
47 |
+
steps=28,
|
48 |
+
progress=gr.Progress()
|
49 |
+
):
|
50 |
+
try:
|
51 |
+
progress(0, desc="Initializing...")
|
52 |
+
generator = torch.Generator(device="cuda").manual_seed(seed)
|
53 |
+
|
54 |
+
# Auto-add trigger words
|
55 |
+
if "super realism" not in prompt.lower():
|
56 |
+
prompt = f"Super Realism, {prompt}"
|
57 |
+
|
58 |
+
# Create callback for progress updates
|
59 |
+
def update_progress(step, _, __):
|
60 |
+
progress((step + 1) / steps, desc="Generating image...")
|
61 |
+
|
62 |
+
# Generate image
|
63 |
+
with torch.inference_mode():
|
64 |
+
image = pipe(
|
65 |
+
prompt=prompt,
|
66 |
+
width=width,
|
67 |
+
height=height,
|
68 |
+
guidance_scale=guidance_scale,
|
69 |
+
num_inference_steps=steps,
|
70 |
+
generator=generator,
|
71 |
+
callback=update_progress
|
72 |
+
).images[0]
|
73 |
+
|
74 |
+
# Safety check
|
75 |
+
progress(1, desc="Safety checking...")
|
76 |
+
safety_input = feature_extractor(image, return_tensors="pt")
|
77 |
+
np_image = np.array(image)
|
78 |
+
safety_result = safety_checker(
|
79 |
+
images=[np_image],
|
80 |
+
clip_input=safety_input.pixel_values
|
81 |
+
)
|
82 |
+
|
83 |
+
if safety_result.nsfw[0]:
|
84 |
+
return Image.new("RGB", (512, 512)), "NSFW content detected"
|
85 |
+
|
86 |
+
return image, "Generation successful"
|
87 |
+
|
88 |
+
except Exception as e:
|
89 |
+
return Image.new("RGB", (512, 512)), f"Error: {str(e)}"
|
90 |
|
91 |
+
# Create Gradio interface with rate limiting
|
92 |
+
with gr.Blocks() as app:
|
93 |
+
gr.Markdown("# Flux Super Realism Generator")
|
94 |
+
|
95 |
+
with gr.Row():
|
96 |
+
with gr.Column():
|
97 |
+
prompt = gr.Textbox(label="Prompt", value="A portrait of a person")
|
98 |
+
seed = gr.Slider(0, 1000, value=42, label="Seed")
|
99 |
+
width = gr.Slider(512, 2048, value=1024, label="Width")
|
100 |
+
height = gr.Slider(512, 2048, value=1024, label="Height")
|
101 |
+
guidance = gr.Slider(1, 20, value=6, label="Guidance Scale")
|
102 |
+
steps = gr.Slider(10, 100, value=28, label="Steps")
|
103 |
+
submit = gr.Button("Generate")
|
104 |
+
|
105 |
+
with gr.Column():
|
106 |
+
output_image = gr.Image(label="Result", type="pil")
|
107 |
+
status = gr.Textbox(label="Status")
|
108 |
+
|
109 |
+
submit.click(
|
110 |
+
generate_image,
|
111 |
+
inputs=[prompt, seed, width, height, guidance, steps],
|
112 |
+
outputs=[output_image, status]
|
113 |
)
|
114 |
+
|
115 |
+
# Rate limiting example (1 request every 30 seconds)
|
116 |
+
app.queue(concurrency_count=1, max_size=3).launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
117 |
|
118 |
+
# For multiple GPU support (advanced)
|
119 |
+
# pipe.enable_model_cpu_offload()
|
120 |
+
# pipe.enable_sequential_cpu_offload()
|