Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -4,15 +4,19 @@ import random
|
|
4 |
import torch
|
5 |
import spaces
|
6 |
from PIL import Image
|
7 |
-
|
8 |
from diffusers import QwenImageEditPipeline
|
9 |
from diffusers.utils import is_xformers_available
|
10 |
-
|
11 |
import os
|
12 |
import base64
|
13 |
import json
|
14 |
from huggingface_hub import InferenceClient
|
15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
def get_caption_language(prompt):
|
17 |
"""Detects if the prompt contains Chinese characters."""
|
18 |
ranges = [
|
@@ -22,7 +26,6 @@ def get_caption_language(prompt):
|
|
22 |
if any(start <= char <= end for start, end in ranges):
|
23 |
return 'zh'
|
24 |
return 'en'
|
25 |
-
|
26 |
def polish_prompt(original_prompt, system_prompt, hf_token):
|
27 |
"""
|
28 |
Rewrites the prompt using a Hugging Face InferenceClient.
|
@@ -31,7 +34,6 @@ def polish_prompt(original_prompt, system_prompt, hf_token):
|
|
31 |
if not hf_token or not hf_token.strip():
|
32 |
gr.Warning("HF Token is required for prompt rewriting but was not provided!")
|
33 |
return original_prompt
|
34 |
-
|
35 |
client = InferenceClient(
|
36 |
provider="cerebras",
|
37 |
api_key=hf_token,
|
@@ -53,7 +55,6 @@ def polish_prompt(original_prompt, system_prompt, hf_token):
|
|
53 |
print(f"Error during Hugging Face API call: {e}")
|
54 |
gr.Warning("Failed to rewrite prompt. Using original.")
|
55 |
return original_prompt
|
56 |
-
|
57 |
SYSTEM_PROMPT_EDIT = '''
|
58 |
# Edit Instruction Rewriter
|
59 |
You are a professional edit instruction rewriter. Your task is to generate a precise, concise, and visually achievable instruction based on the user's intent and the input image.
|
@@ -85,7 +86,6 @@ Please provide the rewritten instruction in a clean `json` format as:
|
|
85 |
"Rewritten": "..."
|
86 |
}
|
87 |
'''
|
88 |
-
|
89 |
dtype = torch.bfloat16
|
90 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
91 |
pipe = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", torch_dtype=dtype).to(device)
|
@@ -94,12 +94,10 @@ pipe.load_lora_weights(
|
|
94 |
"lightx2v/Qwen-Image-Lightning", weight_name="Qwen-Image-Lightning-8steps-V1.1.safetensors"
|
95 |
)
|
96 |
pipe.fuse_lora()
|
97 |
-
|
98 |
if is_xformers_available():
|
99 |
pipe.enable_xformers_memory_efficient_attention()
|
100 |
else:
|
101 |
print("xformers not available or failed to load.")
|
102 |
-
|
103 |
@spaces.GPU(duration=60)
|
104 |
def infer(
|
105 |
image,
|
@@ -116,21 +114,49 @@ def infer(
|
|
116 |
"""
|
117 |
Requires user-provided HF token for prompt rewriting.
|
118 |
"""
|
|
|
119 |
negative_prompt = " "
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
if randomize_seed:
|
121 |
seed = random.randint(0, MAX_SEED)
|
122 |
generator = torch.Generator(device=device).manual_seed(seed)
|
123 |
|
124 |
-
if rewrite_prompt:
|
125 |
-
lang = get_caption_language(prompt)
|
126 |
-
system_prompt = SYSTEM_PROMPT_EDIT
|
127 |
-
polished_prompt = polish_prompt(prompt, system_prompt, hf_token)
|
128 |
-
print(f"Rewritten Prompt: {polished_prompt}")
|
129 |
-
prompt = polished_prompt
|
130 |
-
|
131 |
edited_images = pipe(
|
132 |
image,
|
133 |
-
prompt=
|
134 |
negative_prompt=negative_prompt,
|
135 |
num_inference_steps=num_inference_steps,
|
136 |
generator=generator,
|
@@ -138,7 +164,7 @@ def infer(
|
|
138 |
num_images_per_prompt=num_images_per_prompt,
|
139 |
).images
|
140 |
|
141 |
-
return edited_images, seed
|
142 |
|
143 |
MAX_SEED = np.iinfo(np.int32).max
|
144 |
examples = [
|
@@ -154,7 +180,7 @@ with gr.Blocks() as demo:
|
|
154 |
gr.Markdown("✨ **8-step lightning inferencing with lightx2v's LoRA.**")
|
155 |
gr.Markdown("⚠️ **Prompt rewriting requires your own [Hugging Face token](https://huggingface.co/settings/tokens)**")
|
156 |
gr.Markdown("🚧 **Work in progress, further improvements coming soon.**")
|
157 |
-
|
158 |
with gr.Row():
|
159 |
with gr.Column():
|
160 |
input_image = gr.Image(label="Input Image", type="pil")
|
@@ -168,7 +194,6 @@ with gr.Blocks() as demo:
|
|
168 |
value=0
|
169 |
)
|
170 |
randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
|
171 |
-
|
172 |
with gr.Row():
|
173 |
true_guidance_scale = gr.Slider(
|
174 |
label="True Guidance Scale",
|
@@ -191,12 +216,12 @@ with gr.Blocks() as demo:
|
|
191 |
step=1,
|
192 |
value=1
|
193 |
)
|
194 |
-
|
195 |
-
|
196 |
run_button = gr.Button("Edit", variant="primary")
|
197 |
-
|
198 |
with gr.Column():
|
199 |
result = gr.Gallery(label="Output Images", show_label=False, columns=1)
|
|
|
|
|
200 |
|
201 |
with gr.Group():
|
202 |
rewrite_toggle = gr.Checkbox(label="Use Prompt Rewriter (Requires HF Token)", value=False, interactive=True)
|
@@ -207,7 +232,6 @@ with gr.Blocks() as demo:
|
|
207 |
visible=False,
|
208 |
info="Required for prompt rewriting - get yours from [Hugging Face settings](https://huggingface.co/settings/tokens). API tokens are kept safe locally, but as good practice please make sure to double check the source code. Invalid or missing keys will revert to the original prompt entered."
|
209 |
)
|
210 |
-
|
211 |
def toggle_token_visibility(checked):
|
212 |
return gr.update(visible=checked)
|
213 |
|
@@ -217,8 +241,6 @@ with gr.Blocks() as demo:
|
|
217 |
outputs=[hf_token_input]
|
218 |
)
|
219 |
|
220 |
-
|
221 |
-
|
222 |
gr.on(
|
223 |
triggers=[run_button.click, prompt.submit],
|
224 |
fn=infer,
|
@@ -233,7 +255,24 @@ with gr.Blocks() as demo:
|
|
233 |
hf_token_input,
|
234 |
num_images_per_prompt
|
235 |
],
|
236 |
-
outputs=[result, seed]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
237 |
)
|
238 |
|
239 |
if __name__ == "__main__":
|
|
|
4 |
import torch
|
5 |
import spaces
|
6 |
from PIL import Image
|
|
|
7 |
from diffusers import QwenImageEditPipeline
|
8 |
from diffusers.utils import is_xformers_available
|
|
|
9 |
import os
|
10 |
import base64
|
11 |
import json
|
12 |
from huggingface_hub import InferenceClient
|
13 |
+
import logging
|
14 |
+
#############################
|
15 |
+
os.environ.setdefault('GRADIO_ANALYTICS_ENABLED', 'False')
|
16 |
+
os.environ.setdefault('HF_HUB_DISABLE_TELEMETRY', '1')
|
17 |
+
logging.basicConfig(level=logging.DEBUG)
|
18 |
+
logger = logging.getLogger(__name__)
|
19 |
+
#############################
|
20 |
def get_caption_language(prompt):
|
21 |
"""Detects if the prompt contains Chinese characters."""
|
22 |
ranges = [
|
|
|
26 |
if any(start <= char <= end for start, end in ranges):
|
27 |
return 'zh'
|
28 |
return 'en'
|
|
|
29 |
def polish_prompt(original_prompt, system_prompt, hf_token):
|
30 |
"""
|
31 |
Rewrites the prompt using a Hugging Face InferenceClient.
|
|
|
34 |
if not hf_token or not hf_token.strip():
|
35 |
gr.Warning("HF Token is required for prompt rewriting but was not provided!")
|
36 |
return original_prompt
|
|
|
37 |
client = InferenceClient(
|
38 |
provider="cerebras",
|
39 |
api_key=hf_token,
|
|
|
55 |
print(f"Error during Hugging Face API call: {e}")
|
56 |
gr.Warning("Failed to rewrite prompt. Using original.")
|
57 |
return original_prompt
|
|
|
58 |
SYSTEM_PROMPT_EDIT = '''
|
59 |
# Edit Instruction Rewriter
|
60 |
You are a professional edit instruction rewriter. Your task is to generate a precise, concise, and visually achievable instruction based on the user's intent and the input image.
|
|
|
86 |
"Rewritten": "..."
|
87 |
}
|
88 |
'''
|
|
|
89 |
dtype = torch.bfloat16
|
90 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
91 |
pipe = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", torch_dtype=dtype).to(device)
|
|
|
94 |
"lightx2v/Qwen-Image-Lightning", weight_name="Qwen-Image-Lightning-8steps-V1.1.safetensors"
|
95 |
)
|
96 |
pipe.fuse_lora()
|
|
|
97 |
if is_xformers_available():
|
98 |
pipe.enable_xformers_memory_efficient_attention()
|
99 |
else:
|
100 |
print("xformers not available or failed to load.")
|
|
|
101 |
@spaces.GPU(duration=60)
|
102 |
def infer(
|
103 |
image,
|
|
|
114 |
"""
|
115 |
Requires user-provided HF token for prompt rewriting.
|
116 |
"""
|
117 |
+
original_prompt = prompt # Save original prompt for display
|
118 |
negative_prompt = " "
|
119 |
+
prompt_info = "" # Initialize info text
|
120 |
+
|
121 |
+
# Handle prompt rewriting with status messages
|
122 |
+
if rewrite_prompt:
|
123 |
+
if not hf_token.strip():
|
124 |
+
gr.Warning("HF Token is required for prompt rewriting but was not provided!")
|
125 |
+
prompt_info = f"""## ⚠️ Prompt Rewriting Skipped (No HF Token)
|
126 |
+
**Original Prompt:**
|
127 |
+
{original_prompt}"""
|
128 |
+
rewritten_prompt = original_prompt
|
129 |
+
else:
|
130 |
+
try:
|
131 |
+
rewritten_prompt = polish_prompt(original_prompt, SYSTEM_PROMPT_EDIT, hf_token)
|
132 |
+
prompt_info = f"""## ✅ Prompt Rewrite Successful
|
133 |
+
**Original Prompt:**
|
134 |
+
{original_prompt}
|
135 |
+
|
136 |
+
**Enhanced Prompt:**
|
137 |
+
{rewritten_prompt}"""
|
138 |
+
except Exception as e:
|
139 |
+
gr.Warning(f"Prompt rewriting failed: {str(e)}")
|
140 |
+
rewritten_prompt = original_prompt
|
141 |
+
prompt_info = f"""## ❌ Prompt Rewrite Failed
|
142 |
+
**Original Prompt:**
|
143 |
+
{original_prompt}
|
144 |
+
**Error:**
|
145 |
+
{str(e)}"""
|
146 |
+
else:
|
147 |
+
rewritten_prompt = original_prompt
|
148 |
+
prompt_info = f"""## Original Prompt (No Rewrite)
|
149 |
+
**User Input:**
|
150 |
+
{original_prompt}"""
|
151 |
+
|
152 |
+
# Generate images
|
153 |
if randomize_seed:
|
154 |
seed = random.randint(0, MAX_SEED)
|
155 |
generator = torch.Generator(device=device).manual_seed(seed)
|
156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
edited_images = pipe(
|
158 |
image,
|
159 |
+
prompt=rewritten_prompt,
|
160 |
negative_prompt=negative_prompt,
|
161 |
num_inference_steps=num_inference_steps,
|
162 |
generator=generator,
|
|
|
164 |
num_images_per_prompt=num_images_per_prompt,
|
165 |
).images
|
166 |
|
167 |
+
return edited_images, seed, prompt_info
|
168 |
|
169 |
MAX_SEED = np.iinfo(np.int32).max
|
170 |
examples = [
|
|
|
180 |
gr.Markdown("✨ **8-step lightning inferencing with lightx2v's LoRA.**")
|
181 |
gr.Markdown("⚠️ **Prompt rewriting requires your own [Hugging Face token](https://huggingface.co/settings/tokens)**")
|
182 |
gr.Markdown("🚧 **Work in progress, further improvements coming soon.**")
|
183 |
+
|
184 |
with gr.Row():
|
185 |
with gr.Column():
|
186 |
input_image = gr.Image(label="Input Image", type="pil")
|
|
|
194 |
value=0
|
195 |
)
|
196 |
randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
|
|
|
197 |
with gr.Row():
|
198 |
true_guidance_scale = gr.Slider(
|
199 |
label="True Guidance Scale",
|
|
|
216 |
step=1,
|
217 |
value=1
|
218 |
)
|
|
|
|
|
219 |
run_button = gr.Button("Edit", variant="primary")
|
220 |
+
|
221 |
with gr.Column():
|
222 |
result = gr.Gallery(label="Output Images", show_label=False, columns=1)
|
223 |
+
# New prompt display component
|
224 |
+
prompt_info = gr.Markdown("## Prompt Details", visible=False)
|
225 |
|
226 |
with gr.Group():
|
227 |
rewrite_toggle = gr.Checkbox(label="Use Prompt Rewriter (Requires HF Token)", value=False, interactive=True)
|
|
|
232 |
visible=False,
|
233 |
info="Required for prompt rewriting - get yours from [Hugging Face settings](https://huggingface.co/settings/tokens). API tokens are kept safe locally, but as good practice please make sure to double check the source code. Invalid or missing keys will revert to the original prompt entered."
|
234 |
)
|
|
|
235 |
def toggle_token_visibility(checked):
|
236 |
return gr.update(visible=checked)
|
237 |
|
|
|
241 |
outputs=[hf_token_input]
|
242 |
)
|
243 |
|
|
|
|
|
244 |
gr.on(
|
245 |
triggers=[run_button.click, prompt.submit],
|
246 |
fn=infer,
|
|
|
255 |
hf_token_input,
|
256 |
num_images_per_prompt
|
257 |
],
|
258 |
+
outputs=[result, seed, prompt_info]
|
259 |
+
)
|
260 |
+
|
261 |
+
# Show prompt info box after processing
|
262 |
+
def set_prompt_visible():
|
263 |
+
return gr.update(visible=True)
|
264 |
+
|
265 |
+
run_button.click(
|
266 |
+
fn=set_prompt_visible,
|
267 |
+
inputs=None,
|
268 |
+
outputs=[prompt_info],
|
269 |
+
queue=False
|
270 |
+
)
|
271 |
+
prompt.submit(
|
272 |
+
fn=set_prompt_visible,
|
273 |
+
inputs=None,
|
274 |
+
outputs=[prompt_info],
|
275 |
+
queue=False
|
276 |
)
|
277 |
|
278 |
if __name__ == "__main__":
|