Update gradio_tabs/img_edit.py

gradio_tabs/img_edit.py

@@ -37,111 +37,69 @@ labels_v = [
 ]
 
 
-@torch.compiler.allow_in_graph
 def load_image(img, size):
 	img = Image.open(img).convert('RGB')
 	w, h = img.size
 	img = img.resize((size, size))
 	img = np.asarray(img)
-	img = np.copy(img)
 	img = np.transpose(img, (2, 0, 1))	# 3 x 256 x 256
 
 	return img / 255.0, w, h
 
 
-@torch.compiler.allow_in_graph
 def img_preprocessing(img_path, size):
-	img, w, h = load_image(img_path, size)	# [0, 1]
+	img, w, h = load_image(img_path, size)  # [0, 1]
 	img = torch.from_numpy(img).unsqueeze(0).float()  # [0, 1]
 	imgs_norm = (img - 0.5) * 2.0  # [-1, 1]
 
 	return imgs_norm, w, h
 
 
-# Pre-compile resize transforms for better performance
-resize_transform_cache = {}
-
-def get_resize_transform(size):
-	"""Get cached resize transform - creates once, reuses many times"""
-	if size not in resize_transform_cache:
-		# Only create the transform if it doesn't exist in cache
-		resize_transform_cache[size] = torchvision.transforms.Resize(
-			size,
-			interpolation=torchvision.transforms.InterpolationMode.BILINEAR,
-			antialias=True
-		)
-	return resize_transform_cache[size]
-
-
 def resize(img, size):
-	"""Use cached resize transform"""
-	transform = get_resize_transform((size, size))
+	transform = torchvision.transforms.Compose([
+		torchvision.transforms.Resize((size,size), antialias=True),
+	])
+
 	return transform(img)
 
 
 def resize_back(img, w, h):
-	"""Use cached resize transform for back operation"""
-	transform = get_resize_transform((h, w))
+	transform = torchvision.transforms.Compose([
+		torchvision.transforms.Resize((h, w), antialias=True),
+	])
+
 	return transform(img)
 
 
 def img_denorm(img):
-	img = img.clamp(-1, 1)
+	img = img.clamp(-1, 1).cpu()
 	img = (img - img.min()) / (img.max() - img.min())
 
 	return img
 
 
-def img_postprocessing(img, w, h):
+def img_postprocessing(image, w, h):
 
-	img = resize_back(img, w, h)
-	img = img_denorm(img)
-	img = img.squeeze(0).permute(1, 2, 0).contiguous()	# contiguous() for fast transfer
-	img_output = (img.cpu().numpy() * 255).astype(np.uint8)
+	image = resize_back(image, w, h)
+	image = image.permute(0, 2, 3, 1)
+	edited_image = img_denorm(image)
+	img_output = (edited_image[0].numpy() * 255).astype(np.uint8)
 
-	return img_output	
+	with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as temp_file:
+		imageio.imwrite(temp_file.name, img_output, quality=8)
+		return temp_file.name
 
 
 def img_edit(gen, device):
 
-	@torch.compile
-	def compiled_enc_img(image_tensor, selected_s):
-		"""Compiled version of just the model inference"""
-		return gen.enc_img(image_tensor, labels_v, selected_s)
-
-	@torch.compile
-	def compiled_dec_img(z_s2r, alpha_r2s, feat_rgb):
-		"""Compiled version of just the model inference"""
-		return gen.dec_img(z_s2r, alpha_r2s, feat_rgb)
-
-
-	# Pre-warm the compiled model with dummy data to reduce first-run compilation time
-	def _warmup_model():
-		"""Pre-warm the model compilation with representative shapes"""
-		print("[img_edit] Pre-warming model compilation...")
-		dummy_image = torch.randn(1, 3, 512, 512, device=device)
-		dummy_selected_s = [0.0] * len(labels_v)
-
-		try:
-			with torch.inference_mode():
-				z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(dummy_image, dummy_selected_s)
-				_ = compiled_dec_img(z_s2r, alpha_r2s, feat_rgb)
-			print("[img_edit] Model pre-warming completed successfully")
-		except Exception as e:
-			print(f"[img_edit] Model pre-warming failed (will compile on first use): {e}")
-
-	# Pre-warm the model
-	_warmup_model()
-
 	@spaces.GPU
-	@torch.inference_mode()
+	@torch.no_grad()
 	def edit_img(image, *selected_s):
 
 		image_tensor, w, h = img_preprocessing(image, 512)
 		image_tensor = image_tensor.to(device)
 
-		z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(image_tensor, selected_s)
-		edited_image_tensor = compiled_dec_img(z_s2r, alpha_r2s, feat_rgb)
+		edited_image_tensor = gen.edit_img(image_tensor, labels_v, selected_s)
 
 		# de-norm
 		edited_image = img_postprocessing(edited_image_tensor, w, h)
@@ -178,10 +136,10 @@ def img_edit(gen, device):
 				with gr.Row():
 					with gr.Column(scale=1):
 						with gr.Row():	# Buttons now within a single Row
-							#edit_btn = gr.Button("Edit")
+							edit_btn = gr.Button("Edit")
 							clear_btn = gr.Button("Clear")
-						#with gr.Row():
-						#	animate_btn = gr.Button("Generate")
+						with gr.Row():
+							animate_btn = gr.Button("Generate")
 
 
 
@@ -192,7 +150,7 @@ def img_edit(gen, device):
 						image_output = gr.Image(label="Output Image", type='numpy', interactive=False, width=512)
 
 
-				with gr.Accordion("Control Panel - Using Sliders to Edit Image", open=True):
+				with gr.Accordion("Control Panel", open=True):
 					with gr.Tab("Head"):
 						with gr.Row():
 							for k in labels_k[:3]:
@@ -223,18 +181,12 @@ def img_edit(gen, device):
 								slider = gr.Slider(minimum=-0.2, maximum=0.2, value=0, label=k)
 								inputs_s.append(slider)
 
-		for slider in inputs_s:
-			slider.change(
+
+		edit_btn.click(
 			fn=edit_img,
 			inputs=[image_input] + inputs_s,
 			outputs=[image_output],
-
-			show_progress='hidden',
-
-			trigger_mode='always_last',
-
-			# currently we have a latency around 450ms
-			stream_every=0.5
+			show_progress=True
 		)
 
 		clear_btn.click(