Face detection on CPU

Faster than re-initializing the model on CUDA at each request (300ms vs. 3s)
Should also avoid ONNX runtime errors happening from time to time

app.py

@@ -47,11 +47,12 @@ pipe = StableDiffusionPipeline.from_pretrained(
 ip_model = IPAdapterFaceID(pipe, ip_ckpt, device)
 ip_model_plus = IPAdapterFaceIDPlus(pipe, image_encoder_path, ip_plus_ckpt, device)
 
+app = FaceAnalysis(name="buffalo_l", providers=['CPUExecutionProvider'])
+app.prepare(ctx_id=0, det_size=(640, 640))
+
 @spaces.GPU(enable_queue=True)
 def generate_image(images, prompt, negative_prompt, preserve_face_structure, face_strength, likeness_strength, nfaa_negative_prompt, progress=gr.Progress(track_tqdm=True)):
     pipe.to(device)
-    app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
-    app.prepare(ctx_id=0, det_size=(640, 640))
     
     faceid_all_embeds = []
     first_iteration = True