Update app.py

app.py

@@ -15,7 +15,7 @@ from controlnet_aux import CannyDetector
 from PIL import Image
 import cv2
 
-
+FPS = 24
 dtype = torch.bfloat16 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
@@ -185,10 +185,9 @@ def process_input_video(reference_video, width, height):
         processed_video = process_video_for_canny(video, width, height)
         
         # Create a preview video file for display
-        fps = 24
         with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_file:
             preview_path = tmp_file.name
-            export_to_video(processed_video, preview_path, fps=fps)
+            export_to_video(processed_video, preview_path, fps=FPS)
         
         return preview_path
         
@@ -242,9 +241,8 @@ def generate_video(
         if randomize_seed:
             seed = random.randint(0, 2**32 - 1)
 
-        # Calculate number of frames from duration (24 fps)
-        fps = 24
-        num_frames = int(duration * fps) + 1  # +1 for proper frame count
+        # Calculate number of frames from duration
+        num_frames = int(duration * FPS) + 1  # +1 for proper frame count
         # Ensure num_frames is valid for the model (multiple of temporal compression + 1)
         temporal_compression = pipeline.vae_temporal_compression_ratio
         num_frames = ((num_frames - 1) // temporal_compression) * temporal_compression + 1
@@ -305,7 +303,7 @@ def generate_video(
         progress(0.8, desc="Final denoising and processing...")
         
         # 3. Denoise the upscaled video
-        video_output = pipeline(
+        final_video_frames_np = pipeline(
             prompt=prompt,
             negative_prompt=negative_prompt,
             width=upscaled_width,
@@ -319,22 +317,24 @@ def generate_video(
             decode_noise_scale = 0.025,
             image_cond_noise_scale=0.025,
             generator=torch.Generator(device="cuda").manual_seed(seed),
-            output_type="pil",
+            output_type="np",
         ).frames[0]
 
         progress(0.9, desc="Finalizing output...")
         
-        # 4. Downscale to expected resolution
-        video_output = [frame.resize((width, height)) for frame in video_output]
 
         # Export to temporary file
-        with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_file:
-            output_path = tmp_file.name
-            export_to_video(video_output, output_path, fps=fps)
+        video_uint8_frames = [(frame * 255).astype(np.uint8) for frame in final_video_frames_np]
+        output_filename = "output.mp4"
+        with imageio.get_writer(output_filename, fps=FPS, quality=8, macro_block_size=1) as writer: 
+             for frame_idx, frame_data in enumerate(video_uint8_frames):
+                progress((frame_idx + 1) / len(video_uint8_frames), desc="Encoding video frames...")
+                writer.append_data(frame_data)
+
         
         progress(1.0, desc="Complete!")
         
-        return output_path, seed
+        return output_filename, seed
         
     except Exception as e:
         print(e)
@@ -346,8 +346,9 @@ with gr.Blocks(theme=gr.themes.Ocean(font=[gr.themes.GoogleFont("Lexend Deca"),
         """
         # Canny Control LTX Video Distilled
         
-        LTX Video 0.9.7 Distilled with [control canny ICLoRA](https://huggingface.co/Lightricks/LTX-Video-ICLoRA-canny-13b-0.9.7): 
-        control AI video generation - by concatenation of control signals and Canny LoRA trained on just a few samples  ✨
+        **Fast & canny-controlled video generation using LTX Video 0.9.7 Distilled with [ICLoRA]**(https://huggingface.co/Lightricks/LTX-Video-ICLoRA-canny-13b-0.9.7)
+ 
+        achieved by concatenation of control signals and Canny LoRA trained on just a few samples  ✨
         """
     )