1
commited on
Commit
·
7b7e62e
1
Parent(s):
65dcc19
Update app.py
Browse files
app.py
CHANGED
@@ -14,14 +14,34 @@
|
|
14 |
|
15 |
import dataclasses
|
16 |
import json
|
|
|
|
|
17 |
from pathlib import Path
|
18 |
|
19 |
import gradio as gr
|
20 |
import torch
|
21 |
import spaces
|
|
|
|
|
|
|
22 |
|
23 |
from uno.flux.pipeline import UNOPipeline
|
24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
def get_examples(examples_dir: str = "assets/examples") -> list:
|
26 |
examples = Path(examples_dir)
|
27 |
ans = []
|
@@ -54,6 +74,7 @@ def create_demo(
|
|
54 |
device: str = "cuda" if torch.cuda.is_available() else "cpu",
|
55 |
offload: bool = False,
|
56 |
):
|
|
|
57 |
pipeline = UNOPipeline(model_type, device, offload, only_lora=True, lora_rank=512)
|
58 |
pipeline.gradio_generate = spaces.GPU(duratioin=120)(pipeline.gradio_generate)
|
59 |
|
@@ -229,11 +250,117 @@ def create_demo(
|
|
229 |
],
|
230 |
)
|
231 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
232 |
return demo
|
233 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
234 |
if __name__ == "__main__":
|
235 |
from typing import Literal
|
236 |
-
|
237 |
from transformers import HfArgumentParser
|
238 |
|
239 |
@dataclasses.dataclass
|
@@ -245,10 +372,17 @@ if __name__ == "__main__":
|
|
245 |
metadata={"help": "If True, sequantial offload the models(ae, dit, text encoder) to CPU if not used."}
|
246 |
)
|
247 |
port: int = 7860
|
|
|
248 |
|
249 |
parser = HfArgumentParser([AppArgs])
|
250 |
-
args_tuple = parser.parse_args_into_dataclasses()
|
251 |
args = args_tuple[0]
|
252 |
-
|
|
|
253 |
demo = create_demo(args.name, args.device, args.offload)
|
254 |
-
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
import dataclasses
|
16 |
import json
|
17 |
+
import base64
|
18 |
+
import io
|
19 |
from pathlib import Path
|
20 |
|
21 |
import gradio as gr
|
22 |
import torch
|
23 |
import spaces
|
24 |
+
from PIL import Image as PILImage
|
25 |
+
from fastapi import FastAPI, Body
|
26 |
+
from fastapi.middleware.cors import CORSMiddleware
|
27 |
|
28 |
from uno.flux.pipeline import UNOPipeline
|
29 |
|
30 |
+
# 创建FastAPI应用
|
31 |
+
app = FastAPI()
|
32 |
+
|
33 |
+
# 添加CORS中间件允许跨域请求
|
34 |
+
app.add_middleware(
|
35 |
+
CORSMiddleware,
|
36 |
+
allow_origins=["*"],
|
37 |
+
allow_credentials=True,
|
38 |
+
allow_methods=["*"],
|
39 |
+
allow_headers=["*"],
|
40 |
+
)
|
41 |
+
|
42 |
+
# 设置全局pipeline变量
|
43 |
+
pipeline = None
|
44 |
+
|
45 |
def get_examples(examples_dir: str = "assets/examples") -> list:
|
46 |
examples = Path(examples_dir)
|
47 |
ans = []
|
|
|
74 |
device: str = "cuda" if torch.cuda.is_available() else "cpu",
|
75 |
offload: bool = False,
|
76 |
):
|
77 |
+
global pipeline
|
78 |
pipeline = UNOPipeline(model_type, device, offload, only_lora=True, lora_rank=512)
|
79 |
pipeline.gradio_generate = spaces.GPU(duratioin=120)(pipeline.gradio_generate)
|
80 |
|
|
|
250 |
],
|
251 |
)
|
252 |
|
253 |
+
# 添加API文档
|
254 |
+
with gr.Accordion("API Documentation", open=False):
|
255 |
+
gr.Markdown("""
|
256 |
+
### API Usage
|
257 |
+
|
258 |
+
You can use the following endpoint to generate images programmatically:
|
259 |
+
|
260 |
+
**Endpoint:** `/api/generate`
|
261 |
+
|
262 |
+
**Method:** POST
|
263 |
+
|
264 |
+
**Request Body:**
|
265 |
+
```json
|
266 |
+
{
|
267 |
+
"prompt": "your text prompt",
|
268 |
+
"image_refs": ["base64_encoded_image1", "base64_encoded_image2", ...],
|
269 |
+
"width": 512,
|
270 |
+
"height": 512,
|
271 |
+
"guidance": 4.0,
|
272 |
+
"num_steps": 25,
|
273 |
+
"seed": -1
|
274 |
+
}
|
275 |
+
```
|
276 |
+
|
277 |
+
**Response:**
|
278 |
+
```json
|
279 |
+
{
|
280 |
+
"image": "base64_encoded_generated_image"
|
281 |
+
}
|
282 |
+
```
|
283 |
+
|
284 |
+
**Example JavaScript Usage:**
|
285 |
+
```javascript
|
286 |
+
async function generateImage() {
|
287 |
+
const response = await fetch('/api/generate', {
|
288 |
+
method: 'POST',
|
289 |
+
headers: {
|
290 |
+
'Content-Type': 'application/json',
|
291 |
+
},
|
292 |
+
body: JSON.stringify({
|
293 |
+
prompt: "handsome woman in the city",
|
294 |
+
image_refs: [],
|
295 |
+
width: 512,
|
296 |
+
height: 512
|
297 |
+
}),
|
298 |
+
});
|
299 |
+
|
300 |
+
const data = await response.json();
|
301 |
+
const imgElement = document.getElementById('generatedImage');
|
302 |
+
imgElement.src = `data:image/png;base64,${data.image}`;
|
303 |
+
}
|
304 |
+
```
|
305 |
+
""")
|
306 |
+
|
307 |
return demo
|
308 |
|
309 |
+
# 创建API端点
|
310 |
+
@app.post("/api/generate")
|
311 |
+
async def generate_image(
|
312 |
+
prompt: str = Body(...),
|
313 |
+
width: int = Body(512),
|
314 |
+
height: int = Body(512),
|
315 |
+
guidance: float = Body(4.0),
|
316 |
+
num_steps: int = Body(25),
|
317 |
+
seed: int = Body(-1),
|
318 |
+
image_refs: list = Body([])
|
319 |
+
):
|
320 |
+
global pipeline
|
321 |
+
# 处理参考图像
|
322 |
+
ref_images = []
|
323 |
+
for i in range(min(4, len(image_refs))):
|
324 |
+
if image_refs[i]:
|
325 |
+
try:
|
326 |
+
# 解码base64图像
|
327 |
+
if isinstance(image_refs[i], str) and "base64" in image_refs[i]:
|
328 |
+
# 移除数据URL前缀
|
329 |
+
if "," in image_refs[i]:
|
330 |
+
img_data = image_refs[i].split(",")[1]
|
331 |
+
else:
|
332 |
+
img_data = image_refs[i]
|
333 |
+
|
334 |
+
img_data = base64.b64decode(img_data)
|
335 |
+
ref_img = PILImage.open(io.BytesIO(img_data))
|
336 |
+
ref_images.append(ref_img)
|
337 |
+
else:
|
338 |
+
ref_images.append(None)
|
339 |
+
except:
|
340 |
+
ref_images.append(None)
|
341 |
+
else:
|
342 |
+
ref_images.append(None)
|
343 |
+
|
344 |
+
# 填充至4张图像
|
345 |
+
while len(ref_images) < 4:
|
346 |
+
ref_images.append(None)
|
347 |
+
|
348 |
+
# 调用模型生成图像
|
349 |
+
result_image, _ = pipeline.gradio_generate(
|
350 |
+
prompt, width, height, guidance, num_steps, seed,
|
351 |
+
ref_images[0], ref_images[1], ref_images[2], ref_images[3]
|
352 |
+
)
|
353 |
+
|
354 |
+
# 将结果图像编码为base64
|
355 |
+
buffered = io.BytesIO()
|
356 |
+
result_image.save(buffered, format="PNG")
|
357 |
+
img_str = base64.b64encode(buffered.getvalue()).decode()
|
358 |
+
|
359 |
+
return {"image": img_str}
|
360 |
+
|
361 |
if __name__ == "__main__":
|
362 |
from typing import Literal
|
363 |
+
import uvicorn
|
364 |
from transformers import HfArgumentParser
|
365 |
|
366 |
@dataclasses.dataclass
|
|
|
372 |
metadata={"help": "If True, sequantial offload the models(ae, dit, text encoder) to CPU if not used."}
|
373 |
)
|
374 |
port: int = 7860
|
375 |
+
host: str = "0.0.0.0"
|
376 |
|
377 |
parser = HfArgumentParser([AppArgs])
|
378 |
+
args_tuple = parser.parse_args_into_dataclasses() # type: tuple[AppArgs]
|
379 |
args = args_tuple[0]
|
380 |
+
|
381 |
+
# 创建Gradio demo
|
382 |
demo = create_demo(args.name, args.device, args.offload)
|
383 |
+
|
384 |
+
# 挂载Gradio接口到FastAPI应用
|
385 |
+
app = gr.mount_gradio_app(app, demo, path="/")
|
386 |
+
|
387 |
+
# 使用uvicorn启动FastAPI应用
|
388 |
+
uvicorn.run(app, host=args.host, port=args.port)
|