Spaces:
Runtime error
Runtime error
Commit
•
77ca2e1
1
Parent(s):
461d656
Suggested UI and ZeroGPU compatibility
Browse files
app.py
CHANGED
@@ -1,5 +1,6 @@
|
|
1 |
import gradio as gr
|
2 |
import torch
|
|
|
3 |
from PIL import Image
|
4 |
import os
|
5 |
from transformers import CLIPTokenizer, CLIPTextModel, AutoProcessor, T5EncoderModel, T5TokenizerFast
|
@@ -10,11 +11,10 @@ import torch.nn as nn
|
|
10 |
import math
|
11 |
import logging
|
12 |
import sys
|
13 |
-
from huggingface_hub import snapshot_download
|
14 |
from qwen2_vl.modeling_qwen2_vl import Qwen2VLSimplifiedModel
|
15 |
-
import
|
16 |
|
17 |
-
#
|
18 |
logging.basicConfig(
|
19 |
level=logging.INFO,
|
20 |
format='%(asctime)s - %(levelname)s - %(message)s',
|
@@ -24,10 +24,28 @@ logger = logging.getLogger(__name__)
|
|
24 |
|
25 |
MODEL_ID = "Djrango/Qwen2vl-Flux"
|
26 |
MODEL_CACHE_DIR = "model_cache"
|
27 |
-
|
28 |
-
|
29 |
|
30 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
if not os.path.exists(MODEL_CACHE_DIR):
|
32 |
logger.info("Starting model download...")
|
33 |
try:
|
@@ -41,68 +59,70 @@ if not os.path.exists(MODEL_CACHE_DIR):
|
|
41 |
logger.error(f"Error downloading models: {str(e)}")
|
42 |
raise
|
43 |
|
44 |
-
#
|
45 |
-
logger.info("
|
|
|
|
|
46 |
tokenizer = CLIPTokenizer.from_pretrained(os.path.join(MODEL_CACHE_DIR, "flux/tokenizer"))
|
47 |
text_encoder = CLIPTextModel.from_pretrained(
|
48 |
os.path.join(MODEL_CACHE_DIR, "flux/text_encoder")
|
49 |
-
).to(
|
50 |
|
51 |
text_encoder_two = T5EncoderModel.from_pretrained(
|
52 |
os.path.join(MODEL_CACHE_DIR, "flux/text_encoder_2")
|
53 |
-
).to(
|
54 |
|
55 |
tokenizer_two = T5TokenizerFast.from_pretrained(
|
56 |
-
os.path.join(MODEL_CACHE_DIR, "flux/tokenizer_2")
|
|
|
57 |
|
58 |
-
#
|
59 |
-
logger.info("Loading large models to CPU...")
|
60 |
vae = AutoencoderKL.from_pretrained(
|
61 |
os.path.join(MODEL_CACHE_DIR, "flux/vae")
|
62 |
-
).to(
|
63 |
|
64 |
transformer = FluxTransformer2DModel.from_pretrained(
|
65 |
os.path.join(MODEL_CACHE_DIR, "flux/transformer")
|
66 |
-
).to(
|
67 |
|
68 |
scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
|
69 |
os.path.join(MODEL_CACHE_DIR, "flux/scheduler"),
|
70 |
shift=1
|
71 |
)
|
72 |
|
|
|
73 |
qwen2vl = Qwen2VLSimplifiedModel.from_pretrained(
|
74 |
os.path.join(MODEL_CACHE_DIR, "qwen2-vl")
|
75 |
-
).to(
|
76 |
-
|
77 |
-
qwen2vl_processor = AutoProcessor.from_pretrained(
|
78 |
-
MODEL_ID,
|
79 |
-
subfolder="qwen2-vl",
|
80 |
-
min_pixels=256*28*28,
|
81 |
-
max_pixels=256*28*28
|
82 |
-
)
|
83 |
-
|
84 |
-
# 加载 connector 和 embedder 到 CPU
|
85 |
-
class Qwen2Connector(nn.Module):
|
86 |
-
def __init__(self, input_dim=3584, output_dim=4096):
|
87 |
-
super().__init__()
|
88 |
-
self.linear = nn.Linear(input_dim, output_dim)
|
89 |
-
|
90 |
-
def forward(self, x):
|
91 |
-
return self.linear(x)
|
92 |
|
93 |
-
connector
|
|
|
94 |
connector_path = os.path.join(MODEL_CACHE_DIR, "qwen2-vl/connector.pt")
|
95 |
connector_state = torch.load(connector_path, map_location='cpu')
|
96 |
-
connector_state = {k.replace('module.', ''): v.to(
|
97 |
connector.load_state_dict(connector_state)
|
98 |
|
99 |
-
t5_context_embedder = nn.Linear(4096, 3072).to(
|
100 |
t5_embedder_path = os.path.join(MODEL_CACHE_DIR, "qwen2-vl/t5_embedder.pt")
|
101 |
t5_embedder_state = torch.load(t5_embedder_path, map_location='cpu')
|
102 |
-
t5_embedder_state = {k: v.to(
|
103 |
t5_context_embedder.load_state_dict(t5_embedder_state)
|
104 |
|
105 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
106 |
pipeline = FluxPipeline(
|
107 |
transformer=transformer,
|
108 |
scheduler=scheduler,
|
@@ -111,30 +131,14 @@ pipeline = FluxPipeline(
|
|
111 |
tokenizer=tokenizer,
|
112 |
)
|
113 |
|
114 |
-
# 设置所有模型为eval模式
|
115 |
-
for model in [text_encoder, text_encoder_two, vae, transformer, qwen2vl,
|
116 |
-
connector, t5_context_embedder]:
|
117 |
-
model.requires_grad_(False)
|
118 |
-
model.eval()
|
119 |
-
|
120 |
-
# Aspect ratio options
|
121 |
-
ASPECT_RATIOS = {
|
122 |
-
"1:1": (1024, 1024),
|
123 |
-
"16:9": (1344, 768),
|
124 |
-
"9:16": (768, 1344),
|
125 |
-
"2.4:1": (1536, 640),
|
126 |
-
"3:4": (896, 1152),
|
127 |
-
"4:3": (1152, 896),
|
128 |
-
}
|
129 |
-
|
130 |
def process_image(image):
|
131 |
"""Process image with Qwen2VL model"""
|
132 |
try:
|
133 |
-
#
|
134 |
logger.info("Moving Qwen2VL models to GPU...")
|
135 |
-
qwen2vl.to(
|
136 |
-
connector.to(
|
137 |
-
|
138 |
message = [
|
139 |
{
|
140 |
"role": "user",
|
@@ -156,27 +160,42 @@ def process_image(image):
|
|
156 |
images=[image],
|
157 |
padding=True,
|
158 |
return_tensors="pt"
|
159 |
-
).to(
|
160 |
|
161 |
output_hidden_state, image_token_mask, image_grid_thw = qwen2vl(**inputs)
|
162 |
image_hidden_state = output_hidden_state[image_token_mask].view(1, -1, output_hidden_state.size(-1))
|
163 |
image_hidden_state = connector(image_hidden_state)
|
164 |
|
165 |
-
# 保存结果到 CPU
|
166 |
result = (image_hidden_state.cpu(), image_grid_thw)
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
return result
|
175 |
|
176 |
except Exception as e:
|
177 |
logger.error(f"Error in process_image: {str(e)}")
|
178 |
raise
|
179 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
180 |
def compute_t5_text_embeddings(prompt):
|
181 |
"""Compute T5 embeddings for text prompt"""
|
182 |
if prompt == "":
|
@@ -188,21 +207,15 @@ def compute_t5_text_embeddings(prompt):
|
|
188 |
max_length=256,
|
189 |
truncation=True,
|
190 |
return_tensors="pt"
|
191 |
-
).to(
|
192 |
|
193 |
prompt_embeds = text_encoder_two(text_inputs.input_ids)[0]
|
194 |
-
|
195 |
-
# 将 t5_context_embedder 移到 GPU
|
196 |
-
t5_context_embedder.to(device)
|
197 |
-
prompt_embeds = t5_context_embedder(prompt_embeds)
|
198 |
-
|
199 |
-
# 将 t5_context_embedder 移回 CPU
|
200 |
t5_context_embedder.cpu()
|
201 |
|
202 |
return prompt_embeds
|
203 |
|
204 |
def compute_text_embeddings(prompt=""):
|
205 |
-
"""Compute text embeddings for the prompt"""
|
206 |
with torch.no_grad():
|
207 |
text_inputs = tokenizer(
|
208 |
prompt,
|
@@ -210,18 +223,17 @@ def compute_text_embeddings(prompt=""):
|
|
210 |
max_length=77,
|
211 |
truncation=True,
|
212 |
return_tensors="pt"
|
213 |
-
).to(
|
214 |
|
215 |
prompt_embeds = text_encoder(
|
216 |
text_inputs.input_ids,
|
217 |
output_hidden_states=False
|
218 |
)
|
219 |
-
|
|
|
220 |
|
221 |
-
@spaces.GPU(duration=
|
222 |
-
def
|
223 |
-
num_inference_steps=28, num_images=1, seed=None, aspect_ratio="1:1"):
|
224 |
-
"""Generate images using the pipeline"""
|
225 |
try:
|
226 |
logger.info(f"Starting generation with prompt: {prompt}")
|
227 |
|
@@ -233,31 +245,34 @@ def generate_images(input_image, prompt="", guidance_scale=3.5,
|
|
233 |
logger.info(f"Set random seed to: {seed}")
|
234 |
|
235 |
# Process image with Qwen2VL
|
|
|
236 |
qwen2_hidden_state, image_grid_thw = process_image(input_image)
|
|
|
237 |
|
238 |
# Compute text embeddings
|
|
|
239 |
pooled_prompt_embeds = compute_text_embeddings(prompt)
|
240 |
t5_prompt_embeds = compute_t5_text_embeddings(prompt)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
241 |
|
242 |
# Get dimensions
|
243 |
width, height = ASPECT_RATIOS[aspect_ratio]
|
244 |
logger.info(f"Using dimensions: {width}x{height}")
|
245 |
|
246 |
-
# Generate images
|
247 |
try:
|
248 |
logger.info("Starting image generation...")
|
249 |
-
|
250 |
-
# 将 Transformer 和 VAE 移到 GPU
|
251 |
-
logger.info("Moving Transformer and VAE to GPU...")
|
252 |
-
transformer.to(device)
|
253 |
-
vae.to(device)
|
254 |
-
|
255 |
-
# 更新 pipeline 中的模型引用
|
256 |
-
pipeline.transformer = transformer
|
257 |
-
pipeline.vae = vae
|
258 |
-
|
259 |
output_images = pipeline(
|
260 |
-
prompt_embeds=qwen2_hidden_state.to(
|
261 |
pooled_prompt_embeds=pooled_prompt_embeds,
|
262 |
t5_prompt_embeds=t5_prompt_embeds.repeat(num_images, 1, 1) if t5_prompt_embeds is not None else None,
|
263 |
num_inference_steps=num_inference_steps,
|
@@ -265,15 +280,8 @@ def generate_images(input_image, prompt="", guidance_scale=3.5,
|
|
265 |
height=height,
|
266 |
width=width,
|
267 |
).images
|
268 |
-
|
269 |
logger.info("Image generation completed")
|
270 |
|
271 |
-
# 将 Transformer 和 VAE 移回 CPU
|
272 |
-
logger.info("Moving models back to CPU...")
|
273 |
-
#transformer.cpu()
|
274 |
-
#vae.cpu()
|
275 |
-
torch.cuda.empty_cache()
|
276 |
-
|
277 |
return output_images
|
278 |
|
279 |
except Exception as e:
|
@@ -287,19 +295,32 @@ def generate_images(input_image, prompt="", guidance_scale=3.5,
|
|
287 |
with gr.Blocks(
|
288 |
theme=gr.themes.Soft(),
|
289 |
css="""
|
290 |
-
.container {
|
291 |
-
|
292 |
-
|
293 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
294 |
"""
|
295 |
) as demo:
|
296 |
with gr.Column(elem_classes="container"):
|
297 |
-
gr.Markdown(
|
298 |
-
|
299 |
-
|
300 |
-
|
301 |
-
</div>
|
302 |
-
""")
|
303 |
|
304 |
with gr.Row(equal_height=True):
|
305 |
with gr.Column(scale=1):
|
@@ -309,14 +330,13 @@ with gr.Blocks(
|
|
309 |
height=384,
|
310 |
sources=["upload", "clipboard"]
|
311 |
)
|
312 |
-
|
|
|
|
|
|
|
|
|
313 |
with gr.Accordion("Advanced Settings", open=False):
|
314 |
with gr.Group():
|
315 |
-
prompt = gr.Textbox(
|
316 |
-
label="Text Prompt (Optional)",
|
317 |
-
placeholder="As Long As Possible...",
|
318 |
-
lines=3
|
319 |
-
)
|
320 |
|
321 |
with gr.Row(elem_classes="param-row"):
|
322 |
guidance = gr.Slider(
|
@@ -324,38 +344,48 @@ with gr.Blocks(
|
|
324 |
maximum=10,
|
325 |
value=3.5,
|
326 |
step=0.5,
|
327 |
-
label="Guidance Scale"
|
|
|
328 |
)
|
329 |
steps = gr.Slider(
|
330 |
minimum=1,
|
331 |
-
maximum=
|
332 |
value=28,
|
333 |
step=1,
|
334 |
-
label="Sampling Steps"
|
|
|
335 |
)
|
336 |
|
337 |
with gr.Row(elem_classes="param-row"):
|
338 |
num_images = gr.Slider(
|
339 |
minimum=1,
|
340 |
-
maximum=
|
341 |
-
value=1,
|
342 |
step=1,
|
343 |
-
label="Number of Images"
|
|
|
344 |
)
|
345 |
seed = gr.Number(
|
346 |
label="Random Seed",
|
347 |
value=None,
|
348 |
-
precision=0
|
|
|
349 |
)
|
350 |
aspect_ratio = gr.Radio(
|
351 |
label="Aspect Ratio",
|
352 |
choices=["1:1", "16:9", "9:16", "2.4:1", "3:4", "4:3"],
|
353 |
-
value="1:1"
|
|
|
354 |
)
|
355 |
|
356 |
-
submit_btn = gr.Button(
|
|
|
|
|
|
|
|
|
357 |
|
358 |
with gr.Column(scale=1):
|
|
|
359 |
output_gallery = gr.Gallery(
|
360 |
label="Generated Variations",
|
361 |
columns=2,
|
@@ -363,11 +393,23 @@ with gr.Blocks(
|
|
363 |
height=700,
|
364 |
object_fit="contain",
|
365 |
show_label=True,
|
366 |
-
allow_preview=True
|
|
|
367 |
)
|
|
|
368 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
369 |
submit_btn.click(
|
370 |
-
fn=
|
371 |
inputs=[
|
372 |
input_image,
|
373 |
prompt,
|
@@ -376,14 +418,15 @@ with gr.Blocks(
|
|
376 |
num_images,
|
377 |
seed,
|
378 |
aspect_ratio
|
379 |
-
],
|
380 |
outputs=[output_gallery],
|
381 |
show_progress=True
|
382 |
)
|
383 |
|
|
|
384 |
if __name__ == "__main__":
|
385 |
demo.launch(
|
386 |
-
server_name="0.0.0.0",
|
387 |
-
server_port=7860,
|
388 |
-
share=False
|
389 |
)
|
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
+
import spaces
|
4 |
from PIL import Image
|
5 |
import os
|
6 |
from transformers import CLIPTokenizer, CLIPTextModel, AutoProcessor, T5EncoderModel, T5TokenizerFast
|
|
|
11 |
import math
|
12 |
import logging
|
13 |
import sys
|
|
|
14 |
from qwen2_vl.modeling_qwen2_vl import Qwen2VLSimplifiedModel
|
15 |
+
from huggingface_hub import snapshot_download
|
16 |
|
17 |
+
# Set up logging
|
18 |
logging.basicConfig(
|
19 |
level=logging.INFO,
|
20 |
format='%(asctime)s - %(levelname)s - %(message)s',
|
|
|
24 |
|
25 |
MODEL_ID = "Djrango/Qwen2vl-Flux"
|
26 |
MODEL_CACHE_DIR = "model_cache"
|
27 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
28 |
+
DTYPE = torch.bfloat16
|
29 |
|
30 |
+
# Aspect ratio options
|
31 |
+
ASPECT_RATIOS = {
|
32 |
+
"1:1": (1024, 1024),
|
33 |
+
"16:9": (1344, 768),
|
34 |
+
"9:16": (768, 1344),
|
35 |
+
"2.4:1": (1536, 640),
|
36 |
+
"3:4": (896, 1152),
|
37 |
+
"4:3": (1152, 896),
|
38 |
+
}
|
39 |
+
|
40 |
+
class Qwen2Connector(nn.Module):
|
41 |
+
def __init__(self, input_dim=3584, output_dim=4096):
|
42 |
+
super().__init__()
|
43 |
+
self.linear = nn.Linear(input_dim, output_dim)
|
44 |
+
|
45 |
+
def forward(self, x):
|
46 |
+
return self.linear(x)
|
47 |
+
|
48 |
+
# Download models if not present
|
49 |
if not os.path.exists(MODEL_CACHE_DIR):
|
50 |
logger.info("Starting model download...")
|
51 |
try:
|
|
|
59 |
logger.error(f"Error downloading models: {str(e)}")
|
60 |
raise
|
61 |
|
62 |
+
# Initialize models in global context
|
63 |
+
logger.info("Starting model loading...")
|
64 |
+
|
65 |
+
# Load smaller models to GPU
|
66 |
tokenizer = CLIPTokenizer.from_pretrained(os.path.join(MODEL_CACHE_DIR, "flux/tokenizer"))
|
67 |
text_encoder = CLIPTextModel.from_pretrained(
|
68 |
os.path.join(MODEL_CACHE_DIR, "flux/text_encoder")
|
69 |
+
).to(DTYPE).to(DEVICE)
|
70 |
|
71 |
text_encoder_two = T5EncoderModel.from_pretrained(
|
72 |
os.path.join(MODEL_CACHE_DIR, "flux/text_encoder_2")
|
73 |
+
).to(DTYPE).to(DEVICE)
|
74 |
|
75 |
tokenizer_two = T5TokenizerFast.from_pretrained(
|
76 |
+
os.path.join(MODEL_CACHE_DIR, "flux/tokenizer_2")
|
77 |
+
)
|
78 |
|
79 |
+
# Load larger models to CPU
|
|
|
80 |
vae = AutoencoderKL.from_pretrained(
|
81 |
os.path.join(MODEL_CACHE_DIR, "flux/vae")
|
82 |
+
).to(DTYPE).cpu()
|
83 |
|
84 |
transformer = FluxTransformer2DModel.from_pretrained(
|
85 |
os.path.join(MODEL_CACHE_DIR, "flux/transformer")
|
86 |
+
).to(DTYPE).cpu()
|
87 |
|
88 |
scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
|
89 |
os.path.join(MODEL_CACHE_DIR, "flux/scheduler"),
|
90 |
shift=1
|
91 |
)
|
92 |
|
93 |
+
# Load Qwen2VL to CPU
|
94 |
qwen2vl = Qwen2VLSimplifiedModel.from_pretrained(
|
95 |
os.path.join(MODEL_CACHE_DIR, "qwen2-vl")
|
96 |
+
).to(DTYPE).cpu()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
|
98 |
+
# Load connector and embedder
|
99 |
+
connector = Qwen2Connector().to(DTYPE).cpu()
|
100 |
connector_path = os.path.join(MODEL_CACHE_DIR, "qwen2-vl/connector.pt")
|
101 |
connector_state = torch.load(connector_path, map_location='cpu')
|
102 |
+
connector_state = {k.replace('module.', ''): v.to(DTYPE) for k, v in connector_state.items()}
|
103 |
connector.load_state_dict(connector_state)
|
104 |
|
105 |
+
t5_context_embedder = nn.Linear(4096, 3072).to(DTYPE).cpu()
|
106 |
t5_embedder_path = os.path.join(MODEL_CACHE_DIR, "qwen2-vl/t5_embedder.pt")
|
107 |
t5_embedder_state = torch.load(t5_embedder_path, map_location='cpu')
|
108 |
+
t5_embedder_state = {k: v.to(DTYPE) for k, v in t5_embedder_state.items()}
|
109 |
t5_context_embedder.load_state_dict(t5_embedder_state)
|
110 |
|
111 |
+
# Set all models to eval mode
|
112 |
+
for model in [text_encoder, text_encoder_two, vae, transformer, qwen2vl, connector, t5_context_embedder]:
|
113 |
+
model.requires_grad_(False)
|
114 |
+
model.eval()
|
115 |
+
|
116 |
+
logger.info("All models loaded successfully")
|
117 |
+
|
118 |
+
# Initialize processors and pipeline
|
119 |
+
qwen2vl_processor = AutoProcessor.from_pretrained(
|
120 |
+
MODEL_ID,
|
121 |
+
subfolder="qwen2-vl",
|
122 |
+
min_pixels=256*28*28,
|
123 |
+
max_pixels=256*28*28
|
124 |
+
)
|
125 |
+
|
126 |
pipeline = FluxPipeline(
|
127 |
transformer=transformer,
|
128 |
scheduler=scheduler,
|
|
|
131 |
tokenizer=tokenizer,
|
132 |
)
|
133 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
134 |
def process_image(image):
|
135 |
"""Process image with Qwen2VL model"""
|
136 |
try:
|
137 |
+
# Move Qwen2VL models to GPU
|
138 |
logger.info("Moving Qwen2VL models to GPU...")
|
139 |
+
qwen2vl.to(DEVICE)
|
140 |
+
connector.to(DEVICE)
|
141 |
+
|
142 |
message = [
|
143 |
{
|
144 |
"role": "user",
|
|
|
160 |
images=[image],
|
161 |
padding=True,
|
162 |
return_tensors="pt"
|
163 |
+
).to(DEVICE)
|
164 |
|
165 |
output_hidden_state, image_token_mask, image_grid_thw = qwen2vl(**inputs)
|
166 |
image_hidden_state = output_hidden_state[image_token_mask].view(1, -1, output_hidden_state.size(-1))
|
167 |
image_hidden_state = connector(image_hidden_state)
|
168 |
|
|
|
169 |
result = (image_hidden_state.cpu(), image_grid_thw)
|
170 |
+
|
171 |
+
# Move models back to CPU
|
172 |
+
qwen2vl.cpu()
|
173 |
+
connector.cpu()
|
174 |
+
torch.cuda.empty_cache()
|
175 |
+
|
176 |
+
return result
|
|
|
177 |
|
178 |
except Exception as e:
|
179 |
logger.error(f"Error in process_image: {str(e)}")
|
180 |
raise
|
181 |
|
182 |
+
def resize_image(img, max_pixels=1050000):
|
183 |
+
if not isinstance(img, Image.Image):
|
184 |
+
img = Image.fromarray(img)
|
185 |
+
|
186 |
+
width, height = img.size
|
187 |
+
num_pixels = width * height
|
188 |
+
|
189 |
+
if num_pixels > max_pixels:
|
190 |
+
scale = math.sqrt(max_pixels / num_pixels)
|
191 |
+
new_width = int(width * scale)
|
192 |
+
new_height = int(height * scale)
|
193 |
+
new_width = new_width - (new_width % 8)
|
194 |
+
new_height = new_height - (new_height % 8)
|
195 |
+
img = img.resize((new_width, new_height), Image.LANCZOS)
|
196 |
+
|
197 |
+
return img
|
198 |
+
|
199 |
def compute_t5_text_embeddings(prompt):
|
200 |
"""Compute T5 embeddings for text prompt"""
|
201 |
if prompt == "":
|
|
|
207 |
max_length=256,
|
208 |
truncation=True,
|
209 |
return_tensors="pt"
|
210 |
+
).to(DEVICE)
|
211 |
|
212 |
prompt_embeds = text_encoder_two(text_inputs.input_ids)[0]
|
213 |
+
prompt_embeds = t5_context_embedder.to(DEVICE)(prompt_embeds)
|
|
|
|
|
|
|
|
|
|
|
214 |
t5_context_embedder.cpu()
|
215 |
|
216 |
return prompt_embeds
|
217 |
|
218 |
def compute_text_embeddings(prompt=""):
|
|
|
219 |
with torch.no_grad():
|
220 |
text_inputs = tokenizer(
|
221 |
prompt,
|
|
|
223 |
max_length=77,
|
224 |
truncation=True,
|
225 |
return_tensors="pt"
|
226 |
+
).to(DEVICE)
|
227 |
|
228 |
prompt_embeds = text_encoder(
|
229 |
text_inputs.input_ids,
|
230 |
output_hidden_states=False
|
231 |
)
|
232 |
+
pooled_prompt_embeds = prompt_embeds.pooler_output
|
233 |
+
return pooled_prompt_embeds
|
234 |
|
235 |
+
@spaces.GPU(duration=75)
|
236 |
+
def generate(input_image, prompt="", guidance_scale=3.5, num_inference_steps=28, num_images=2, seed=None, aspect_ratio="1:1", progress=gr.Progress(track_tqdm=True)):
|
|
|
|
|
237 |
try:
|
238 |
logger.info(f"Starting generation with prompt: {prompt}")
|
239 |
|
|
|
245 |
logger.info(f"Set random seed to: {seed}")
|
246 |
|
247 |
# Process image with Qwen2VL
|
248 |
+
logger.info("Processing input image with Qwen2VL...")
|
249 |
qwen2_hidden_state, image_grid_thw = process_image(input_image)
|
250 |
+
logger.info("Image processing completed")
|
251 |
|
252 |
# Compute text embeddings
|
253 |
+
logger.info("Computing text embeddings...")
|
254 |
pooled_prompt_embeds = compute_text_embeddings(prompt)
|
255 |
t5_prompt_embeds = compute_t5_text_embeddings(prompt)
|
256 |
+
logger.info("Text embeddings computed")
|
257 |
+
|
258 |
+
# Move Transformer and VAE to GPU
|
259 |
+
logger.info("Moving Transformer and VAE to GPU...")
|
260 |
+
transformer.to(DEVICE)
|
261 |
+
vae.to(DEVICE)
|
262 |
+
|
263 |
+
# Update pipeline models
|
264 |
+
pipeline.transformer = transformer
|
265 |
+
pipeline.vae = vae
|
266 |
+
logger.info("Models moved to GPU")
|
267 |
|
268 |
# Get dimensions
|
269 |
width, height = ASPECT_RATIOS[aspect_ratio]
|
270 |
logger.info(f"Using dimensions: {width}x{height}")
|
271 |
|
|
|
272 |
try:
|
273 |
logger.info("Starting image generation...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
274 |
output_images = pipeline(
|
275 |
+
prompt_embeds=qwen2_hidden_state.to(DEVICE).repeat(num_images, 1, 1),
|
276 |
pooled_prompt_embeds=pooled_prompt_embeds,
|
277 |
t5_prompt_embeds=t5_prompt_embeds.repeat(num_images, 1, 1) if t5_prompt_embeds is not None else None,
|
278 |
num_inference_steps=num_inference_steps,
|
|
|
280 |
height=height,
|
281 |
width=width,
|
282 |
).images
|
|
|
283 |
logger.info("Image generation completed")
|
284 |
|
|
|
|
|
|
|
|
|
|
|
|
|
285 |
return output_images
|
286 |
|
287 |
except Exception as e:
|
|
|
295 |
with gr.Blocks(
|
296 |
theme=gr.themes.Soft(),
|
297 |
css="""
|
298 |
+
.container {
|
299 |
+
max-width: 1200px;
|
300 |
+
margin: auto;
|
301 |
+
}
|
302 |
+
.header {
|
303 |
+
text-align: center;
|
304 |
+
margin: 20px 0 40px 0;
|
305 |
+
padding: 20px;
|
306 |
+
background: #f7f7f7;
|
307 |
+
border-radius: 12px;
|
308 |
+
}
|
309 |
+
.param-row {
|
310 |
+
padding: 10px 0;
|
311 |
+
}
|
312 |
+
footer {
|
313 |
+
margin-top: 40px;
|
314 |
+
padding: 20px;
|
315 |
+
border-top: 1px solid #eee;
|
316 |
+
}
|
317 |
"""
|
318 |
) as demo:
|
319 |
with gr.Column(elem_classes="container"):
|
320 |
+
gr.Markdown(
|
321 |
+
"""# 🎨 Qwen2vl-Flux Image Variation Demo
|
322 |
+
Generate creative variations of your images with optional text guidance"""
|
323 |
+
)
|
|
|
|
|
324 |
|
325 |
with gr.Row(equal_height=True):
|
326 |
with gr.Column(scale=1):
|
|
|
330 |
height=384,
|
331 |
sources=["upload", "clipboard"]
|
332 |
)
|
333 |
+
prompt = gr.Textbox(
|
334 |
+
label="Text Prompt (Optional)",
|
335 |
+
placeholder="As Long As Possible...",
|
336 |
+
lines=3
|
337 |
+
)
|
338 |
with gr.Accordion("Advanced Settings", open=False):
|
339 |
with gr.Group():
|
|
|
|
|
|
|
|
|
|
|
340 |
|
341 |
with gr.Row(elem_classes="param-row"):
|
342 |
guidance = gr.Slider(
|
|
|
344 |
maximum=10,
|
345 |
value=3.5,
|
346 |
step=0.5,
|
347 |
+
label="Guidance Scale",
|
348 |
+
info="Higher values follow prompt more closely"
|
349 |
)
|
350 |
steps = gr.Slider(
|
351 |
minimum=1,
|
352 |
+
maximum=50,
|
353 |
value=28,
|
354 |
step=1,
|
355 |
+
label="Sampling Steps",
|
356 |
+
info="More steps = better quality but slower"
|
357 |
)
|
358 |
|
359 |
with gr.Row(elem_classes="param-row"):
|
360 |
num_images = gr.Slider(
|
361 |
minimum=1,
|
362 |
+
maximum=4,
|
363 |
+
value=1,
|
364 |
step=1,
|
365 |
+
label="Number of Images",
|
366 |
+
info="Generate multiple variations at once"
|
367 |
)
|
368 |
seed = gr.Number(
|
369 |
label="Random Seed",
|
370 |
value=None,
|
371 |
+
precision=0,
|
372 |
+
info="Set for reproducible results"
|
373 |
)
|
374 |
aspect_ratio = gr.Radio(
|
375 |
label="Aspect Ratio",
|
376 |
choices=["1:1", "16:9", "9:16", "2.4:1", "3:4", "4:3"],
|
377 |
+
value="1:1",
|
378 |
+
info="Choose aspect ratio for generated images"
|
379 |
)
|
380 |
|
381 |
+
submit_btn = gr.Button(
|
382 |
+
"🎨 Generate Variations",
|
383 |
+
variant="primary",
|
384 |
+
size="lg"
|
385 |
+
)
|
386 |
|
387 |
with gr.Column(scale=1):
|
388 |
+
# Output Section
|
389 |
output_gallery = gr.Gallery(
|
390 |
label="Generated Variations",
|
391 |
columns=2,
|
|
|
393 |
height=700,
|
394 |
object_fit="contain",
|
395 |
show_label=True,
|
396 |
+
allow_preview=True,
|
397 |
+
preview=True
|
398 |
)
|
399 |
+
error_message = gr.Textbox(visible=False)
|
400 |
|
401 |
+
with gr.Row(elem_classes="footer"):
|
402 |
+
gr.Markdown("""
|
403 |
+
### Tips:
|
404 |
+
- 📸 Upload any image to get started
|
405 |
+
- 💡 Add an optional text prompt to guide the generation
|
406 |
+
- 🎯 Adjust guidance scale to control prompt influence
|
407 |
+
- ⚙️ Increase steps for higher quality
|
408 |
+
- 🎲 Use seeds for reproducible results
|
409 |
+
""")
|
410 |
+
|
411 |
submit_btn.click(
|
412 |
+
fn=generate,
|
413 |
inputs=[
|
414 |
input_image,
|
415 |
prompt,
|
|
|
418 |
num_images,
|
419 |
seed,
|
420 |
aspect_ratio
|
421 |
+
],
|
422 |
outputs=[output_gallery],
|
423 |
show_progress=True
|
424 |
)
|
425 |
|
426 |
+
# Launch the app
|
427 |
if __name__ == "__main__":
|
428 |
demo.launch(
|
429 |
+
server_name="0.0.0.0", # Listen on all network interfaces
|
430 |
+
server_port=7860, # Use a specific port
|
431 |
+
share=False, # Disable public URL sharing
|
432 |
)
|