Update gradio_tabs/animation.py

gradio_tabs/animation.py

@@ -36,74 +36,64 @@ labels_v = [
 	13, 24, 17, 26
 ]
 
-@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 vid_preprocessing(vid_path, size):
 	vid_dict = torchvision.io.read_video(vid_path, pts_unit='sec')
-	vid = vid_dict[0].permute(0, 3, 1, 2) # tchw
+	vid = vid_dict[0].permute(0, 3, 1, 2).unsqueeze(0)	# btchw
 	fps = vid_dict[2]['video_fps']
 	vid_norm = (vid / 255.0 - 0.5) * 2.0  # [-1, 1]
-	vid_norm = resize(vid_norm, size) # tchw	
+
+	vid_norm = torch.cat([
+		resize(vid_norm[:, i, :, :, :], size).unsqueeze(1) for i in range(vid.size(1))
+	], dim=1)
 
 	return vid_norm, fps
 
 
 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 vid_denorm(vid):
-	vid = vid.clamp(-1, 1)
+	vid = vid.clamp(-1, 1).cpu()
 	vid = (vid - vid.min()) / (vid.max() - vid.min())
 
 	return vid
@@ -111,30 +101,24 @@ def vid_denorm(vid):
 
 def img_postprocessing(image, w, h):
 
-	img = resize_back(image, w, h)
-
-	# Denormalize ON GPU (avoid early CPU transfer)
-	img = img.clamp(-1, 1)	# Still on GPU
-	img = (img - img.min()) / (img.max() - img.min())  # Still on GPU
+	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)
 
-	# Single optimized CPU transfer
-	img = img.squeeze(0).permute(1, 2, 0).contiguous()	# contiguous() for fast transfer
-	img_output = (img.cpu().numpy() * 255).astype(np.uint8)  # Single CPU transfer
-
-	# return the Numpy array directly, since Gradio supports it
-	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 vid_postprocessing(video, w, h, fps):
-	# video: TCHW
+	# video: BCTHW
 
-	t,c,_,_ = video.size()
-	vid = resize_back(video, w, h)
-	vid = vid_denorm(vid)
-
-	vid = rearrange(vid, "t c h w -> t h w c")	# T H W C
-	vid_np = (vid.cpu().numpy() * 255).astype('uint8')
+	b,c,t,_,_ = video.size()
+	vid_batch = resize_back(rearrange(video, "b c t h w -> (b t) c h w"), w, h)
+	vid = rearrange(vid_batch, "(b t) c h w -> b t h w c", b=b)	# B T H W C
+	vid_np = (vid_denorm(vid[0]).numpy() * 255).astype('uint8')
 
 	with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_file:
 		imageio.mimwrite(temp_file.name, vid_np, fps=fps, codec='libx264', quality=8)
@@ -142,59 +126,15 @@ def vid_postprocessing(video, w, h, fps):
 
 
 def animation(gen, chunk_size, 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)
-
-	@torch.compile
-	def compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target_batch):
-		"""Compiled version of animate_batch for animation tab"""
-		return gen.dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target_batch)
-
-	# 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_video = torch.randn(chunk_size, 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}")
-
-		try:
-			with torch.inference_mode():
-				z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(dummy_image, dummy_selected_s)
-				_ = compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, dummy_video[0], dummy_video)
-			print("[img_animation] Model pre-warming completed successfully")
-		except Exception as e:
-			print(f"[img_animation] 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_media(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)
@@ -202,35 +142,16 @@ def animation(gen, chunk_size, device):
 		return edited_image
 
 	@spaces.GPU
-	@torch.inference_mode()
+	@torch.no_grad()
 	def animate_media(image, video, *selected_s):
 
 		image_tensor, w, h = img_preprocessing(image, 512)
 		vid_target_tensor, fps = vid_preprocessing(video, 512)
 		image_tensor = image_tensor.to(device)
-		video_target_tensor = vid_target_tensor.to(device) #tchw
-		
-		img_start = video_target_tensor[0:1,:,:,:]
-		
-		res = []
-		t, c, h, w = video_target_tensor.size()
-
-		chunks = t // chunk_size
-		if t%chunk_size == 0:
-			vid_target_tensor_batch = torch.zeros(chunk_size * chunks, c, h, w).to(device)
-		else:
-			vid_target_tensor_batch = torch.zeros(chunk_size * (chunks + 1), c, h, w).to(device)
-		vid_target_tensor_batch[:t] = video_target_tensor
-
-		z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(image_tensor, selected_s)
-		for i in range(chunks+1):
-
-			img_target_batch = vid_target_tensor_batch[i * chunk_size:(i + 1) * chunk_size, :, :, :]
-			img_animated_batch = compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target_batch)
-
-			res.append(img_animated_batch)
-		animated_video = torch.cat(res, dim=0)[:t] # TCHW
-		edited_image = animated_video[0:1,:,:,:]
+		video_target_tensor = vid_target_tensor.to(device)
+
+		animated_video = gen.animate_batch(image_tensor, video_target_tensor, labels_v, selected_s, chunk_size)
+		edited_image = animated_video[:,:,0,:,:]
 
 		# postprocessing
 		animated_video = vid_postprocessing(animated_video, w, h, fps)
@@ -241,7 +162,7 @@ def animation(gen, chunk_size, device):
 	def clear_media():
 		return None, None, *([0] * len(labels_k))
 
-	
+    
 	with gr.Tab("Image Animation"):
 
 		inputs_s = []
@@ -281,10 +202,11 @@ def animation(gen, chunk_size, device):
 				with gr.Row():
 					with gr.Column(scale=1):
 						with gr.Row():	# Buttons now within a single Row
-							#edit_btn = gr.Button("Edit", elem_id="button_edit",)
-							animate_btn = gr.Button("Animate", elem_id="button_animate")
-						with gr.Row():
+							edit_btn = gr.Button("Edit", elem_id="button_edit",)
 							clear_btn = gr.Button("Clear", elem_id="button_clear")
+						with gr.Row():
+							animate_btn = gr.Button("Animate", elem_id="button_animate")
+
 
 
 			with gr.Column(scale=1):
@@ -299,7 +221,7 @@ def animation(gen, chunk_size, device):
 						#video_output.render()
 						video_output = gr.Video(label="Output Video", elem_id="output_vid", width=512)#.render()
 
-				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]:
@@ -329,23 +251,20 @@ def animation(gen, chunk_size, device):
 							for k in labels_k[12:14]:
 								slider = gr.Slider(minimum=-0.2, maximum=0.2, value=0, label=k, elem_id="slider_"+str(k))
 								inputs_s.append(slider)
-		
-		for slider in inputs_s:
-			slider.change(
-				fn=edit_media,
-				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
-			)
+
+
+		edit_btn.click(
+			fn=edit_media,
+			inputs=[image_input] + inputs_s,
+			outputs=[image_output],
+			show_progress=True
+		)
 
 		animate_btn.click(
 			fn=animate_media,
 			inputs=[image_input, video_input] + inputs_s,
 			outputs=[image_output, video_output],
-			show_progress=True
+            show_progress=True
 		)
 
 		clear_btn.click(
@@ -361,14 +280,14 @@ def animation(gen, chunk_size, device):
 				['./data/source/macron.png', './data/driving/driving1.mp4', 0.14,0,-0.26,-0.29,-0.11,0,-0.13,-0.18,0,0,0,0,-0.02,0.07],
 				['./data/source/portrait3.png', './data/driving/driving1.mp4', -0.03,0.21,-0.31,-0.12,-0.11,0,-0.05,-0.16,0,0,0,0,-0.02,0.07],
 				['./data/source/einstein.png','./data/driving/driving2.mp4',-0.31,0,0,0.16,0.08,0,-0.07,0,0.13,0,0,0,0,0],
-				['./data/source/portrait1.png', './data/driving/driving4.mp4', 0, 0, -0.17, -0.19, 0.25, 0, 0, -0.086,
+                ['./data/source/portrait1.png', './data/driving/driving4.mp4', 0, 0, -0.17, -0.19, 0.25, 0, 0, -0.086,
 				 0.087, 0, 0, 0, 0, 0],
 				['./data/source/portrait2.png','./data/driving/driving8.mp4',0,0,-0.25,0,0,0,0,0,0,0.126,0,0,0,0],
 				
 			],
-			fn=animate_media,
+            fn=animate_media,
 			inputs=[image_input, video_input] + inputs_s,
-			outputs=[image_output, video_output],
+            outputs=[image_output, video_output],
 		)