Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,75 +1,26 @@
|
|
1 |
-
import
|
2 |
import random
|
3 |
import warnings
|
4 |
-
import
|
5 |
import gradio as gr
|
6 |
import numpy as np
|
7 |
import spaces
|
8 |
import torch
|
9 |
-
import torch.nn as nn
|
10 |
from diffusers import FluxImg2ImgPipeline
|
|
|
11 |
from gradio_imageslider import ImageSlider
|
12 |
from PIL import Image
|
13 |
from huggingface_hub import snapshot_download
|
14 |
import requests
|
15 |
|
16 |
-
#
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
self.conv5 = nn.Conv2d(num_feat + 4 * num_grow_ch, num_feat, 3, 1, 1)
|
25 |
-
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
|
26 |
-
|
27 |
-
def forward(self, x):
|
28 |
-
x1 = self.lrelu(self.conv1(x))
|
29 |
-
x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
|
30 |
-
x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
|
31 |
-
x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
|
32 |
-
x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
|
33 |
-
return x5 * 0.2 + x
|
34 |
-
|
35 |
-
class RRDB(nn.Module):
|
36 |
-
def __init__(self, num_feat, num_grow_ch=32):
|
37 |
-
super(RRDB, self).__init__()
|
38 |
-
self.rdb1 = ResidualDenseBlock(num_feat, num_grow_ch)
|
39 |
-
self.rdb2 = ResidualDenseBlock(num_feat, num_grow_ch)
|
40 |
-
self.rdb3 = ResidualDenseBlock(num_feat, num_grow_ch)
|
41 |
-
|
42 |
-
def forward(self, x):
|
43 |
-
out = self.rdb1(x)
|
44 |
-
out = self.rdb2(out)
|
45 |
-
out = self.rdb3(out)
|
46 |
-
return out * 0.2 + x
|
47 |
-
|
48 |
-
class RRDBNet(nn.Module):
|
49 |
-
def __init__(self, num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4):
|
50 |
-
super(RRDBNet, self).__init__()
|
51 |
-
self.scale = scale
|
52 |
-
|
53 |
-
self.conv_first = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
|
54 |
-
self.body = nn.Sequential(*[RRDB(num_feat, num_grow_ch) for _ in range(num_block)])
|
55 |
-
self.conv_body = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
|
56 |
-
|
57 |
-
# Upsampling
|
58 |
-
self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
|
59 |
-
self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
|
60 |
-
self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
|
61 |
-
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
|
62 |
-
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
|
63 |
-
|
64 |
-
def forward(self, x):
|
65 |
-
fea = self.conv_first(x)
|
66 |
-
trunk = self.conv_body(self.body(fea))
|
67 |
-
fea = fea + trunk
|
68 |
-
|
69 |
-
fea = self.lrelu(self.conv_up1(nn.functional.interpolate(fea, scale_factor=2, mode='nearest')))
|
70 |
-
fea = self.lrelu(self.conv_up2(nn.functional.interpolate(fea, scale_factor=2, mode='nearest')))
|
71 |
-
out = self.conv_last(self.lrelu(self.conv_hr(fea)))
|
72 |
-
return out
|
73 |
|
74 |
css = """
|
75 |
#col-container {
|
@@ -82,10 +33,14 @@ css = """
|
|
82 |
}
|
83 |
"""
|
84 |
|
|
|
|
|
|
|
|
|
85 |
# Get HuggingFace token
|
86 |
huggingface_token = os.getenv("HF_TOKEN")
|
87 |
|
88 |
-
# Download FLUX model
|
89 |
print("π₯ Downloading FLUX model...")
|
90 |
model_path = snapshot_download(
|
91 |
repo_id="black-forest-labs/FLUX.1-dev",
|
@@ -95,276 +50,324 @@ model_path = snapshot_download(
|
|
95 |
token=huggingface_token,
|
96 |
)
|
97 |
|
98 |
-
# Load
|
99 |
-
print("π₯ Loading
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
pipe = FluxImg2ImgPipeline.from_pretrained(
|
101 |
model_path,
|
102 |
-
torch_dtype=torch.
|
103 |
-
use_safetensors=True
|
104 |
)
|
105 |
-
|
106 |
-
# Enable memory optimizations
|
107 |
pipe.enable_vae_tiling()
|
108 |
pipe.enable_vae_slicing()
|
109 |
-
pipe.vae.enable_tiling()
|
110 |
-
pipe.vae.enable_slicing()
|
111 |
-
|
112 |
-
# Download and load ESRGAN 4x-UltraSharp model
|
113 |
-
print("π₯ Loading ESRGAN 4x-UltraSharp...")
|
114 |
-
esrgan_path = "4x-UltraSharp.pth"
|
115 |
-
if not os.path.exists(esrgan_path):
|
116 |
-
print("Downloading ESRGAN model...")
|
117 |
-
url = "https://huggingface.co/uwg/upscaler/resolve/main/ESRGAN/4x-UltraSharp.pth"
|
118 |
-
response = requests.get(url)
|
119 |
-
with open(esrgan_path, "wb") as f:
|
120 |
-
f.write(response.content)
|
121 |
-
|
122 |
-
# Initialize ESRGAN model
|
123 |
-
esrgan_model = RRDBNet(
|
124 |
-
num_in_ch=3,
|
125 |
-
num_out_ch=3,
|
126 |
-
num_feat=64,
|
127 |
-
num_block=23,
|
128 |
-
num_grow_ch=32,
|
129 |
-
scale=4
|
130 |
-
)
|
131 |
-
|
132 |
-
# Load state dict
|
133 |
-
state_dict = torch.load(esrgan_path, map_location='cpu')
|
134 |
-
if 'params_ema' in state_dict:
|
135 |
-
state_dict = state_dict['params_ema']
|
136 |
-
elif 'params' in state_dict:
|
137 |
-
state_dict = state_dict['params']
|
138 |
-
|
139 |
-
# Clean state dict keys if needed
|
140 |
-
cleaned_state_dict = {}
|
141 |
-
for k, v in state_dict.items():
|
142 |
-
if k.startswith('module.'):
|
143 |
-
cleaned_state_dict[k[7:]] = v
|
144 |
-
else:
|
145 |
-
cleaned_state_dict[k] = v
|
146 |
-
|
147 |
-
esrgan_model.load_state_dict(cleaned_state_dict, strict=False)
|
148 |
-
esrgan_model.eval()
|
149 |
|
150 |
print("β
All models loaded successfully!")
|
151 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
MAX_SEED = 1000000
|
153 |
-
|
154 |
|
155 |
|
156 |
-
def
|
157 |
-
"""
|
158 |
-
|
|
|
|
|
159 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
160 |
|
161 |
-
def truncate_prompt(prompt, max_tokens=75):
|
162 |
-
"""Truncate prompt to avoid CLIP token limit (77 tokens)"""
|
163 |
-
if not prompt:
|
164 |
-
return ""
|
165 |
-
|
166 |
-
# Simple truncation by character count (rough approximation)
|
167 |
-
if len(prompt) > 250: # ~75 tokens
|
168 |
-
return prompt[:250] + "..."
|
169 |
-
return prompt
|
170 |
|
|
|
|
|
|
|
|
|
|
|
171 |
|
172 |
-
|
173 |
-
"""Prepare image for processing"""
|
174 |
-
w, h = image.size
|
175 |
-
|
176 |
-
# Limit input size
|
177 |
-
if w > max_size or h > max_size:
|
178 |
-
image.thumbnail((max_size, max_size), Image.LANCZOS)
|
179 |
-
|
180 |
-
return image
|
181 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
182 |
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
with torch.no_grad():
|
201 |
-
output =
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
217 |
|
218 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
219 |
def enhance_image(
|
220 |
-
|
221 |
-
|
222 |
seed,
|
223 |
randomize_seed,
|
224 |
num_inference_steps,
|
225 |
-
denoising_strength,
|
226 |
upscale_factor,
|
|
|
|
|
|
|
227 |
progress=gr.Progress(track_tqdm=True),
|
228 |
):
|
229 |
"""Main enhancement function"""
|
230 |
-
|
231 |
-
raise gr.Error("Please upload an image")
|
232 |
-
|
233 |
-
# Clear memory
|
234 |
-
torch.cuda.empty_cache()
|
235 |
-
gc.collect()
|
236 |
-
|
237 |
try:
|
238 |
-
|
239 |
-
|
240 |
-
|
241 |
-
|
242 |
-
|
243 |
-
prompt = truncate_prompt(prompt.strip() if prompt else "high quality, detailed")
|
244 |
-
|
245 |
-
# Prepare input image
|
246 |
-
input_image = prepare_image(input_image)
|
247 |
-
original_size = input_image.size
|
248 |
-
|
249 |
-
# Step 1: ESRGAN upscale on GPU
|
250 |
-
gr.Info(f"π Upscaling with ESRGAN x{upscale_factor}...")
|
251 |
-
|
252 |
-
# Move ESRGAN to GPU for faster processing
|
253 |
-
esrgan_model.to("cuda")
|
254 |
-
upscaled_image = esrgan_upscale(input_image, esrgan_model, device="cuda", upscale_factor=upscale_factor)
|
255 |
-
|
256 |
-
# Move ESRGAN back to CPU to free memory
|
257 |
-
esrgan_model.to("cpu")
|
258 |
-
torch.cuda.empty_cache()
|
259 |
-
|
260 |
-
# Ensure dimensions are multiples of 16 for FLUX
|
261 |
-
w, h = upscaled_image.size
|
262 |
-
new_w = make_multiple_16(w)
|
263 |
-
new_h = make_multiple_16(h)
|
264 |
-
|
265 |
-
if new_w != w or new_h != h:
|
266 |
-
# Pad image to meet requirements
|
267 |
-
padded = Image.new('RGB', (new_w, new_h))
|
268 |
-
padded.paste(upscaled_image, (0, 0))
|
269 |
-
upscaled_image = padded
|
270 |
-
|
271 |
-
# Step 2: FLUX enhancement
|
272 |
-
gr.Info("π¨ Enhancing with FLUX...")
|
273 |
-
|
274 |
-
# Move pipeline to GPU
|
275 |
-
pipe.to("cuda")
|
276 |
-
|
277 |
-
# Generate with FLUX
|
278 |
-
generator = torch.Generator(device="cuda").manual_seed(seed)
|
279 |
-
|
280 |
-
with torch.inference_mode():
|
281 |
-
result = pipe(
|
282 |
-
prompt=prompt,
|
283 |
-
image=upscaled_image,
|
284 |
-
strength=denoising_strength,
|
285 |
-
num_inference_steps=num_inference_steps,
|
286 |
-
guidance_scale=3.5, # Recommended for FLUX.1-dev to reduce artifacts
|
287 |
-
height=new_h,
|
288 |
-
width=new_w,
|
289 |
-
generator=generator,
|
290 |
-
).images[0]
|
291 |
-
|
292 |
-
# Crop back if we padded
|
293 |
-
if new_w != w or new_h != h:
|
294 |
-
result = result.crop((0, 0, w, h))
|
295 |
-
|
296 |
-
# Move pipeline back to CPU
|
297 |
-
pipe.to("cpu")
|
298 |
-
torch.cuda.empty_cache()
|
299 |
-
gc.collect()
|
300 |
-
|
301 |
-
# Prepare images for slider (before/after)
|
302 |
-
input_resized = input_image.resize(result.size, Image.LANCZOS)
|
303 |
-
|
304 |
-
gr.Info("β
Enhancement complete!")
|
305 |
-
return [input_resized, result], seed
|
306 |
-
|
307 |
except Exception as e:
|
308 |
-
|
309 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
310 |
esrgan_model.to("cpu")
|
311 |
-
|
312 |
-
|
313 |
-
raise gr.Error(f"Enhancement failed: {str(e)}")
|
314 |
|
315 |
|
316 |
# Create Gradio interface
|
317 |
-
with gr.Blocks(css=css) as demo:
|
318 |
gr.HTML("""
|
319 |
<div class="main-header">
|
320 |
-
<h1
|
321 |
-
<p>Upload an image to upscale 2
|
322 |
-
<p>
|
323 |
</div>
|
324 |
-
""")
|
325 |
-
|
326 |
with gr.Row():
|
327 |
with gr.Column(scale=1):
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
333 |
)
|
334 |
|
335 |
-
|
336 |
-
label="
|
337 |
-
placeholder="
|
338 |
-
value="high quality, ultra detailed, sharp",
|
339 |
lines=2
|
340 |
)
|
341 |
|
342 |
-
|
|
|
343 |
upscale_factor = gr.Slider(
|
344 |
-
label="Upscale
|
345 |
-
minimum=
|
346 |
maximum=4,
|
347 |
step=1,
|
348 |
-
value=
|
349 |
-
info="
|
350 |
)
|
351 |
|
352 |
num_inference_steps = gr.Slider(
|
353 |
-
label="
|
354 |
-
minimum=
|
355 |
-
maximum=
|
356 |
step=1,
|
357 |
-
value=
|
358 |
info="More steps = better quality but slower"
|
359 |
)
|
360 |
|
361 |
denoising_strength = gr.Slider(
|
362 |
-
label="
|
363 |
-
minimum=0.
|
364 |
-
maximum=0
|
365 |
step=0.05,
|
366 |
-
value=0.
|
367 |
-
info="
|
368 |
)
|
369 |
|
370 |
with gr.Row():
|
@@ -372,58 +375,124 @@ with gr.Blocks(css=css) as demo:
|
|
372 |
label="Randomize seed",
|
373 |
value=True
|
374 |
)
|
375 |
-
seed = gr.
|
376 |
label="Seed",
|
377 |
-
|
|
|
|
|
|
|
|
|
378 |
)
|
379 |
|
380 |
enhance_btn = gr.Button(
|
381 |
-
"Upscale",
|
382 |
variant="primary",
|
383 |
size="lg"
|
384 |
)
|
385 |
-
|
386 |
-
with gr.Column(scale=2):
|
387 |
-
|
|
|
388 |
result_slider = ImageSlider(
|
389 |
type="pil",
|
390 |
-
|
391 |
-
|
392 |
-
|
393 |
-
|
394 |
-
|
395 |
-
used_seed = gr.Number(
|
396 |
-
label="Seed Used",
|
397 |
-
interactive=False,
|
398 |
-
visible=False
|
399 |
)
|
400 |
-
|
401 |
# Event handler
|
402 |
enhance_btn.click(
|
403 |
fn=enhance_image,
|
404 |
inputs=[
|
405 |
input_image,
|
406 |
-
|
407 |
seed,
|
408 |
randomize_seed,
|
409 |
num_inference_steps,
|
410 |
-
denoising_strength,
|
411 |
upscale_factor,
|
|
|
|
|
|
|
412 |
],
|
413 |
-
outputs=[result_slider
|
414 |
)
|
415 |
|
416 |
gr.HTML("""
|
417 |
-
<div style="margin-top: 2rem;
|
418 |
-
<p
|
419 |
-
<p>β‘ Optimized for ZeroGPU with automatic memory management</p>
|
420 |
-
<p>π Note: User is responsible for obtaining commercial license from Flux Dev if using image commercially under their license.</p>
|
421 |
</div>
|
422 |
""")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
423 |
|
424 |
if __name__ == "__main__":
|
425 |
-
demo.queue(
|
426 |
-
share=False,
|
427 |
-
server_name="0.0.0.0",
|
428 |
-
server_port=7860
|
429 |
-
)
|
|
|
1 |
+
import logging
|
2 |
import random
|
3 |
import warnings
|
4 |
+
import os
|
5 |
import gradio as gr
|
6 |
import numpy as np
|
7 |
import spaces
|
8 |
import torch
|
|
|
9 |
from diffusers import FluxImg2ImgPipeline
|
10 |
+
from transformers import AutoProcessor, AutoModelForCausalLM
|
11 |
from gradio_imageslider import ImageSlider
|
12 |
from PIL import Image
|
13 |
from huggingface_hub import snapshot_download
|
14 |
import requests
|
15 |
|
16 |
+
# For ESRGAN (requires pip install basicsr gfpgan)
|
17 |
+
try:
|
18 |
+
from basicsr.archs.rrdbnet_arch import RRDBNet
|
19 |
+
from basicsr.utils import img2tensor, tensor2img
|
20 |
+
USE_ESRGAN = True
|
21 |
+
except ImportError:
|
22 |
+
USE_ESRGAN = False
|
23 |
+
warnings.warn("basicsr not installed; falling back to LANCZOS interpolation.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
css = """
|
26 |
#col-container {
|
|
|
33 |
}
|
34 |
"""
|
35 |
|
36 |
+
# Device setup - Default to CPU, let runtime handle GPU
|
37 |
+
power_device = "ZeroGPU"
|
38 |
+
device = "cpu"
|
39 |
+
|
40 |
# Get HuggingFace token
|
41 |
huggingface_token = os.getenv("HF_TOKEN")
|
42 |
|
43 |
+
# Download FLUX model
|
44 |
print("π₯ Downloading FLUX model...")
|
45 |
model_path = snapshot_download(
|
46 |
repo_id="black-forest-labs/FLUX.1-dev",
|
|
|
50 |
token=huggingface_token,
|
51 |
)
|
52 |
|
53 |
+
# Load Florence-2 model for image captioning on CPU
|
54 |
+
print("π₯ Loading Florence-2 model...")
|
55 |
+
florence_model = AutoModelForCausalLM.from_pretrained(
|
56 |
+
"microsoft/Florence-2-large",
|
57 |
+
torch_dtype=torch.float32, # Force CPU dtype
|
58 |
+
trust_remote_code=True,
|
59 |
+
attn_implementation="eager"
|
60 |
+
).to(device)
|
61 |
+
florence_processor = AutoProcessor.from_pretrained(
|
62 |
+
"microsoft/Florence-2-large",
|
63 |
+
trust_remote_code=True
|
64 |
+
)
|
65 |
+
|
66 |
+
# Load FLUX Img2Img pipeline on CPU
|
67 |
+
print("π₯ Loading FLUX Img2Img...")
|
68 |
pipe = FluxImg2ImgPipeline.from_pretrained(
|
69 |
model_path,
|
70 |
+
torch_dtype=torch.float32 # Force CPU dtype
|
|
|
71 |
)
|
|
|
|
|
72 |
pipe.enable_vae_tiling()
|
73 |
pipe.enable_vae_slicing()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
|
75 |
print("β
All models loaded successfully!")
|
76 |
|
77 |
+
# Download ESRGAN model if using
|
78 |
+
if USE_ESRGAN:
|
79 |
+
esrgan_path = "4x-UltraSharp.pth"
|
80 |
+
if not os.path.exists(esrgan_path):
|
81 |
+
url = "https://huggingface.co/uwg/upscaler/resolve/main/ESRGAN/4x-UltraSharp.pth"
|
82 |
+
with open(esrgan_path, "wb") as f:
|
83 |
+
f.write(requests.get(url).content)
|
84 |
+
esrgan_model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
|
85 |
+
state_dict = torch.load(esrgan_path)['params_ema']
|
86 |
+
esrgan_model.load_state_dict(state_dict)
|
87 |
+
esrgan_model.eval()
|
88 |
+
|
89 |
MAX_SEED = 1000000
|
90 |
+
MAX_PIXEL_BUDGET = 8192 * 8192 # Increased for tiling support
|
91 |
|
92 |
|
93 |
+
def generate_caption(image):
|
94 |
+
"""Generate detailed caption using Florence-2"""
|
95 |
+
try:
|
96 |
+
task_prompt = "<MORE_DETAILED_CAPTION>"
|
97 |
+
prompt = task_prompt
|
98 |
|
99 |
+
inputs = florence_processor(text=prompt, images=image, return_tensors="pt").to(device)
|
100 |
+
|
101 |
+
generated_ids = florence_model.generate(
|
102 |
+
input_ids=inputs["input_ids"],
|
103 |
+
pixel_values=inputs["pixel_values"],
|
104 |
+
max_new_tokens=1024,
|
105 |
+
num_beams=3,
|
106 |
+
do_sample=True,
|
107 |
+
)
|
108 |
+
|
109 |
+
generated_text = florence_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
|
110 |
+
parsed_answer = florence_processor.post_process_generation(generated_text, task=task_prompt, image_size=(image.width, image.height))
|
111 |
+
|
112 |
+
caption = parsed_answer[task_prompt]
|
113 |
+
return caption
|
114 |
+
except Exception as e:
|
115 |
+
print(f"Caption generation failed: {e}")
|
116 |
+
return "a high quality detailed image"
|
117 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
118 |
|
119 |
+
def process_input(input_image, upscale_factor):
|
120 |
+
"""Process input image and handle size constraints"""
|
121 |
+
w, h = input_image.size
|
122 |
+
w_original, h_original = w, h
|
123 |
+
aspect_ratio = w / h
|
124 |
|
125 |
+
was_resized = False
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
126 |
|
127 |
+
if w * h * upscale_factor**2 > MAX_PIXEL_BUDGET:
|
128 |
+
warnings.warn(
|
129 |
+
f"Requested output image is too large ({w * upscale_factor}x{h * upscale_factor}). Resizing to fit budget."
|
130 |
+
)
|
131 |
+
gr.Info(
|
132 |
+
f"Requested output image is too large. Resizing input to fit within pixel budget."
|
133 |
+
)
|
134 |
+
target_input_pixels = MAX_PIXEL_BUDGET / (upscale_factor ** 2)
|
135 |
+
scale = (target_input_pixels / (w * h)) ** 0.5
|
136 |
+
new_w = int(w * scale) - int(w * scale) % 8
|
137 |
+
new_h = int(h * scale) - int(h * scale) % 8
|
138 |
+
input_image = input_image.resize((new_w, new_h), resample=Image.LANCZOS)
|
139 |
+
was_resized = True
|
140 |
|
141 |
+
return input_image, w_original, h_original, was_resized
|
142 |
+
|
143 |
+
|
144 |
+
def load_image_from_url(url):
|
145 |
+
"""Load image from URL"""
|
146 |
+
try:
|
147 |
+
response = requests.get(url, stream=True)
|
148 |
+
response.raise_for_status()
|
149 |
+
return Image.open(response.raw)
|
150 |
+
except Exception as e:
|
151 |
+
raise gr.Error(f"Failed to load image from URL: {e}")
|
152 |
+
|
153 |
+
|
154 |
+
def esrgan_upscale(image, scale=4):
|
155 |
+
if not USE_ESRGAN:
|
156 |
+
return image.resize((image.width * scale, image.height * scale), resample=Image.LANCZOS)
|
157 |
+
img = img2tensor(np.array(image) / 255., bgr2rgb=False, float32=True)
|
158 |
with torch.no_grad():
|
159 |
+
output = esrgan_model(img.unsqueeze(0)).squeeze()
|
160 |
+
output_img = tensor2img(output, rgb2bgr=False, min_max=(0, 1))
|
161 |
+
return Image.fromarray(output_img)
|
162 |
+
|
163 |
+
|
164 |
+
def tiled_flux_img2img(pipe, prompt, image, strength, steps, guidance, generator, tile_size=1024, overlap=32):
|
165 |
+
"""Tiled Img2Img to mimic Ultimate SD Upscaler tiling"""
|
166 |
+
w, h = image.size
|
167 |
+
output = image.copy() # Start with the control image
|
168 |
+
|
169 |
+
for x in range(0, w, tile_size - overlap):
|
170 |
+
for y in range(0, h, tile_size - overlap):
|
171 |
+
tile_w = min(tile_size, w - x)
|
172 |
+
tile_h = min(tile_size, h - y)
|
173 |
+
tile = image.crop((x, y, x + tile_w, y + tile_h))
|
174 |
|
175 |
+
# Run Flux on tile
|
176 |
+
gen_tile = pipe(
|
177 |
+
prompt=prompt,
|
178 |
+
image=tile,
|
179 |
+
strength=strength,
|
180 |
+
num_inference_steps=steps,
|
181 |
+
guidance_scale=guidance,
|
182 |
+
height=tile_h,
|
183 |
+
width=tile_w,
|
184 |
+
generator=generator,
|
185 |
+
).images[0]
|
186 |
|
187 |
+
# Paste with blending if overlap
|
188 |
+
if overlap > 0:
|
189 |
+
paste_box = (x, y, x + tile_w, y + tile_h)
|
190 |
+
if x > 0 or y > 0:
|
191 |
+
# Simple linear blend on overlaps
|
192 |
+
mask = Image.new('L', (tile_w, tile_h), 255)
|
193 |
+
if x > 0:
|
194 |
+
for i in range(overlap):
|
195 |
+
for j in range(tile_h):
|
196 |
+
mask.putpixel((i, j), int(255 * (i / overlap)))
|
197 |
+
if y > 0:
|
198 |
+
for i in range(tile_w):
|
199 |
+
for j in range(overlap):
|
200 |
+
mask.putpixel((i, j), int(255 * (j / overlap)))
|
201 |
+
output.paste(gen_tile, paste_box, mask)
|
202 |
+
else:
|
203 |
+
output.paste(gen_tile, paste_box)
|
204 |
+
else:
|
205 |
+
output.paste(gen_tile, (x, y))
|
206 |
+
|
207 |
+
return output
|
208 |
+
|
209 |
+
|
210 |
+
@spaces.GPU(duration=120)
|
211 |
def enhance_image(
|
212 |
+
image_input,
|
213 |
+
image_url,
|
214 |
seed,
|
215 |
randomize_seed,
|
216 |
num_inference_steps,
|
|
|
217 |
upscale_factor,
|
218 |
+
denoising_strength,
|
219 |
+
use_generated_caption,
|
220 |
+
custom_prompt,
|
221 |
progress=gr.Progress(track_tqdm=True),
|
222 |
):
|
223 |
"""Main enhancement function"""
|
224 |
+
# Move models to GPU with fallback to CPU
|
|
|
|
|
|
|
|
|
|
|
|
|
225 |
try:
|
226 |
+
device = "cuda"
|
227 |
+
pipe.to(device)
|
228 |
+
florence_model.to(device)
|
229 |
+
if USE_ESRGAN:
|
230 |
+
esrgan_model.to(device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
231 |
except Exception as e:
|
232 |
+
print(f"GPU error: {e}, falling back to CPU")
|
233 |
+
device = "cpu"
|
234 |
+
|
235 |
+
# Handle image input
|
236 |
+
if image_input is not None:
|
237 |
+
input_image = image_input
|
238 |
+
elif image_url:
|
239 |
+
input_image = load_image_from_url(image_url)
|
240 |
+
else:
|
241 |
+
raise gr.Error("Please provide an image (upload or URL)")
|
242 |
+
|
243 |
+
if randomize_seed:
|
244 |
+
seed = random.randint(0, MAX_SEED)
|
245 |
+
|
246 |
+
true_input_image = input_image
|
247 |
+
|
248 |
+
# Process input image
|
249 |
+
input_image, w_original, h_original, was_resized = process_input(
|
250 |
+
input_image, upscale_factor
|
251 |
+
)
|
252 |
+
|
253 |
+
# Generate caption if requested
|
254 |
+
if use_generated_caption:
|
255 |
+
gr.Info("π Generating image caption...")
|
256 |
+
generated_caption = generate_caption(input_image)
|
257 |
+
prompt = generated_caption
|
258 |
+
else:
|
259 |
+
prompt = custom_prompt if custom_prompt.strip() else ""
|
260 |
+
|
261 |
+
generator = torch.Generator(device=device).manual_seed(seed)
|
262 |
+
|
263 |
+
gr.Info("π Upscaling image...")
|
264 |
+
|
265 |
+
# Initial upscale
|
266 |
+
if USE_ESRGAN and upscale_factor == 4:
|
267 |
+
control_image = esrgan_upscale(input_image, upscale_factor)
|
268 |
+
else:
|
269 |
+
w, h = input_image.size
|
270 |
+
control_image = input_image.resize((w * upscale_factor, h * upscale_factor), resample=Image.LANCZOS)
|
271 |
+
|
272 |
+
# Tiled Flux Img2Img for refinement
|
273 |
+
image = tiled_flux_img2img(
|
274 |
+
pipe,
|
275 |
+
prompt,
|
276 |
+
control_image,
|
277 |
+
denoising_strength,
|
278 |
+
num_inference_steps,
|
279 |
+
1.0, # Hardcoded guidance_scale to 1
|
280 |
+
generator,
|
281 |
+
tile_size=1024,
|
282 |
+
overlap=32
|
283 |
+
)
|
284 |
+
|
285 |
+
if was_resized:
|
286 |
+
gr.Info(f"π Resizing output to target size: {w_original * upscale_factor}x{h_original * upscale_factor}")
|
287 |
+
image = image.resize((w_original * upscale_factor, h_original * upscale_factor), resample=Image.LANCZOS)
|
288 |
+
|
289 |
+
# Resize input image to match output size for slider alignment
|
290 |
+
resized_input = true_input_image.resize(image.size, resample=Image.LANCZOS)
|
291 |
+
|
292 |
+
# Move back to CPU to release GPU
|
293 |
+
pipe.to("cpu")
|
294 |
+
florence_model.to("cpu")
|
295 |
+
if USE_ESRGAN:
|
296 |
esrgan_model.to("cpu")
|
297 |
+
|
298 |
+
return [resized_input, image]
|
|
|
299 |
|
300 |
|
301 |
# Create Gradio interface
|
302 |
+
with gr.Blocks(css=css, title="π¨ AI Image Upscaler - Florence-2 + FLUX") as demo:
|
303 |
gr.HTML("""
|
304 |
<div class="main-header">
|
305 |
+
<h1>π¨ AI Image Upscaler</h1>
|
306 |
+
<p>Upload an image or provide a URL to upscale it using Florence-2 captioning and FLUX upscaling</p>
|
307 |
+
<p>Currently running on <strong>{}</strong></p>
|
308 |
</div>
|
309 |
+
""".format(power_device))
|
310 |
+
|
311 |
with gr.Row():
|
312 |
with gr.Column(scale=1):
|
313 |
+
gr.HTML("<h3>π€ Input</h3>")
|
314 |
+
|
315 |
+
with gr.Tabs():
|
316 |
+
with gr.TabItem("π Upload Image"):
|
317 |
+
input_image = gr.Image(
|
318 |
+
label="Upload Image",
|
319 |
+
type="pil",
|
320 |
+
height=200 # Made smaller
|
321 |
+
)
|
322 |
+
|
323 |
+
with gr.TabItem("π Image URL"):
|
324 |
+
image_url = gr.Textbox(
|
325 |
+
label="Image URL",
|
326 |
+
placeholder="https://example.com/image.jpg",
|
327 |
+
value="https://upload.wikimedia.org/wikipedia/commons/thumb/a/a7/Example.jpg/800px-Example.jpg"
|
328 |
+
)
|
329 |
+
|
330 |
+
gr.HTML("<h3>ποΈ Caption Settings</h3>")
|
331 |
+
|
332 |
+
use_generated_caption = gr.Checkbox(
|
333 |
+
label="Use AI-generated caption (Florence-2)",
|
334 |
+
value=True,
|
335 |
+
info="Generate detailed caption automatically"
|
336 |
)
|
337 |
|
338 |
+
custom_prompt = gr.Textbox(
|
339 |
+
label="Custom Prompt (optional)",
|
340 |
+
placeholder="Enter custom prompt or leave empty for generated caption",
|
|
|
341 |
lines=2
|
342 |
)
|
343 |
|
344 |
+
gr.HTML("<h3>βοΈ Upscaling Settings</h3>")
|
345 |
+
|
346 |
upscale_factor = gr.Slider(
|
347 |
+
label="Upscale Factor",
|
348 |
+
minimum=1,
|
349 |
maximum=4,
|
350 |
step=1,
|
351 |
+
value=2,
|
352 |
+
info="How much to upscale the image"
|
353 |
)
|
354 |
|
355 |
num_inference_steps = gr.Slider(
|
356 |
+
label="Number of Inference Steps",
|
357 |
+
minimum=8,
|
358 |
+
maximum=50,
|
359 |
step=1,
|
360 |
+
value=25,
|
361 |
info="More steps = better quality but slower"
|
362 |
)
|
363 |
|
364 |
denoising_strength = gr.Slider(
|
365 |
+
label="Denoising Strength",
|
366 |
+
minimum=0.0,
|
367 |
+
maximum=1.0,
|
368 |
step=0.05,
|
369 |
+
value=0.3,
|
370 |
+
info="Controls how much the image is transformed"
|
371 |
)
|
372 |
|
373 |
with gr.Row():
|
|
|
375 |
label="Randomize seed",
|
376 |
value=True
|
377 |
)
|
378 |
+
seed = gr.Slider(
|
379 |
label="Seed",
|
380 |
+
minimum=0,
|
381 |
+
maximum=MAX_SEED,
|
382 |
+
step=1,
|
383 |
+
value=42,
|
384 |
+
interactive=True
|
385 |
)
|
386 |
|
387 |
enhance_btn = gr.Button(
|
388 |
+
"π Upscale Image",
|
389 |
variant="primary",
|
390 |
size="lg"
|
391 |
)
|
392 |
+
|
393 |
+
with gr.Column(scale=2): # Larger scale for results
|
394 |
+
gr.HTML("<h3>π Results</h3>")
|
395 |
+
|
396 |
result_slider = ImageSlider(
|
397 |
type="pil",
|
398 |
+
interactive=False, # Disable interactivity to prevent uploads
|
399 |
+
height=600, # Made larger
|
400 |
+
elem_id="result_slider",
|
401 |
+
label=None # Remove default label
|
|
|
|
|
|
|
|
|
|
|
402 |
)
|
403 |
+
|
404 |
# Event handler
|
405 |
enhance_btn.click(
|
406 |
fn=enhance_image,
|
407 |
inputs=[
|
408 |
input_image,
|
409 |
+
image_url,
|
410 |
seed,
|
411 |
randomize_seed,
|
412 |
num_inference_steps,
|
|
|
413 |
upscale_factor,
|
414 |
+
denoising_strength,
|
415 |
+
use_generated_caption,
|
416 |
+
custom_prompt,
|
417 |
],
|
418 |
+
outputs=[result_slider]
|
419 |
)
|
420 |
|
421 |
gr.HTML("""
|
422 |
+
<div style="margin-top: 2rem; padding: 1rem; background: #f0f0f0; border-radius: 8px;">
|
423 |
+
<p><strong>Note:</strong> This upscaler uses the Flux dev model. Users are responsible for obtaining commercial rights if used commercially under their license.</p>
|
|
|
|
|
424 |
</div>
|
425 |
""")
|
426 |
+
|
427 |
+
# Custom CSS for slider
|
428 |
+
gr.HTML("""
|
429 |
+
<style>
|
430 |
+
#result_slider .slider {
|
431 |
+
width: 100% !important;
|
432 |
+
max-width: inherit !important;
|
433 |
+
}
|
434 |
+
#result_slider img {
|
435 |
+
object-fit: contain !important;
|
436 |
+
width: 100% !important;
|
437 |
+
height: auto !important;
|
438 |
+
}
|
439 |
+
#result_slider .gr-button-tool {
|
440 |
+
display: none !important;
|
441 |
+
}
|
442 |
+
#result_slider .gr-button-undo {
|
443 |
+
display: none !important;
|
444 |
+
}
|
445 |
+
#result_slider .gr-button-clear {
|
446 |
+
display: none !important;
|
447 |
+
}
|
448 |
+
#result_slider .badge-container .badge {
|
449 |
+
display: none !important;
|
450 |
+
}
|
451 |
+
#result_slider .badge-container::before {
|
452 |
+
content: "Before";
|
453 |
+
position: absolute;
|
454 |
+
top: 10px;
|
455 |
+
left: 10px;
|
456 |
+
background: rgba(0,0,0,0.5);
|
457 |
+
color: white;
|
458 |
+
padding: 5px;
|
459 |
+
border-radius: 5px;
|
460 |
+
z-index: 10;
|
461 |
+
}
|
462 |
+
#result_slider .badge-container::after {
|
463 |
+
content: "After";
|
464 |
+
position: absolute;
|
465 |
+
top: 10px;
|
466 |
+
right: 10px;
|
467 |
+
background: rgba(0,0,0,0.5);
|
468 |
+
color: white;
|
469 |
+
padding: 5px;
|
470 |
+
border-radius: 5px;
|
471 |
+
z-index: 10;
|
472 |
+
}
|
473 |
+
#result_slider .fullscreen img {
|
474 |
+
object-fit: contain !important;
|
475 |
+
width: 100vw !important;
|
476 |
+
height: 100vh !important;
|
477 |
+
position: absolute;
|
478 |
+
top: 0;
|
479 |
+
left: 0;
|
480 |
+
}
|
481 |
+
</style>
|
482 |
+
""")
|
483 |
+
|
484 |
+
# JS to set slider default position to middle
|
485 |
+
gr.HTML("""
|
486 |
+
<script>
|
487 |
+
document.addEventListener('DOMContentLoaded', function() {
|
488 |
+
const sliderInput = document.querySelector('#result_slider input[type="range"]');
|
489 |
+
if (sliderInput) {
|
490 |
+
sliderInput.value = 50;
|
491 |
+
sliderInput.dispatchEvent(new Event('input'));
|
492 |
+
}
|
493 |
+
});
|
494 |
+
</script>
|
495 |
+
""")
|
496 |
|
497 |
if __name__ == "__main__":
|
498 |
+
demo.queue().launch(share=True, server_name="0.0.0.0", server_port=7860)
|
|
|
|
|
|
|
|