experiment

app.py

@@ -185,7 +185,9 @@ def frames_to_mp4_base64(frames, fps = 15):
     
     return "data:video/mp4;base64,"
 
-def initialize_vae_decoder(use_taehv=False, use_trt=False):
+# note: we set use_taehv to be able to use other resolutions
+# this might impact performance
+def initialize_vae_decoder(use_taehv=True, use_trt=False):
     if use_trt:
         from demo_utils.vae import VAETRTWrapper
         print("Initializing TensorRT VAE Decoder...")

app_with_streaming.py

@@ -253,8 +253,6 @@ def video_generation_handler_streaming(prompt, seed=42, fps=15, width=DEFAULT_WI
         seed = random.randint(0, 2**32 - 1)
     
 
-    print(f"🎬 video_generation_handler_streaming called, seed: {seed}, duration: {duration}s, fps: {fps}, width: {width}, height: {height}")
-
     # Setup
     conditional_dict = text_encoder(text_prompts=[prompt])
     for key, value in conditional_dict.items():
@@ -267,6 +265,9 @@ def video_generation_handler_streaming(prompt, seed=42, fps=15, width=DEFAULT_WI
     # Calculate latent dimensions based on actual width/height (assuming 8x downsampling)
     latent_height = height // 8
     latent_width = width // 8
+
+    print(f"🎬 video_generation_handler_streaming called, seed: {seed}, duration: {duration}s, fps: {fps}, width: {width}, height: {height}")
+
     noise = torch.randn([1, 21, 16, latent_height, latent_width], device=gpu, dtype=torch.float16, generator=rnd)
     
     vae_cache, latents_cache = None, None