Update app.py

app.py

@@ -96,11 +96,10 @@ def construct_video_pipeline(model_id: str, lora_path: str):
     pipe.unload_lora_weights()
     
     return pipe
-
-def generate_video(prompt: str, frame1_url: str, frame2_url: str, guidance_scale: float, num_frames: int, num_inference_steps: int) -> bytes:
+def generate_video(prompt: str, frame1_path: str, frame2_path: str, guidance_scale: float, num_frames: int, num_inference_steps: int) -> bytes:
     # Load and preprocess frames
-    cond_frame1 = Image.open(requests.get(frame1_url, stream=True).raw)
-    cond_frame2 = Image.open(requests.get(frame2_url, stream=True).raw)
+    cond_frame1 = Image.open(frame1_path)
+    cond_frame2 = Image.open(frame2_path)
     
     height, width = 720, 1280
     cond_frame1 = resize_image_to_bucket(cond_frame1, bucket_reso=(width, height))
@@ -145,6 +144,54 @@ def generate_video(prompt: str, frame1_url: str, frame2_url: str, guidance_scale
         video_bytes = video_file.read()
     
     return video_bytes
+# def generate_video(prompt: str, frame1_url: str, frame2_url: str, guidance_scale: float, num_frames: int, num_inference_steps: int) -> bytes:
+#     # Load and preprocess frames
+#     cond_frame1 = Image.open(requests.get(frame1_url, stream=True).raw)
+#     cond_frame2 = Image.open(requests.get(frame2_url, stream=True).raw)
+    
+#     height, width = 720, 1280
+#     cond_frame1 = resize_image_to_bucket(cond_frame1, bucket_reso=(width, height))
+#     cond_frame2 = resize_image_to_bucket(cond_frame2, bucket_reso=(width, height))
+    
+#     cond_video = np.zeros(shape=(num_frames, height, width, 3))
+#     cond_video[0], cond_video[-1] = np.array(cond_frame1), np.array(cond_frame2)
+#     cond_video = torch.from_numpy(cond_video.copy()).permute(0, 3, 1, 2)
+#     cond_video = torch.stack([video_transforms(x) for x in cond_video], dim=0).unsqueeze(0)
+    
+#     # Initialize pipeline
+#     model_id = "hunyuanvideo-community/HunyuanVideo"
+#     lora_path = hf_hub_download("dashtoon/hunyuan-video-keyframe-control-lora", "i2v.sft")  # Replace with the actual LORA path
+#     pipe = construct_video_pipeline(model_id, lora_path)
+    
+#     with torch.no_grad():
+#         image_or_video = cond_video.to(device="cuda", dtype=pipe.dtype)
+#         image_or_video = image_or_video.permute(0, 2, 1, 3, 4).contiguous()  # [B, F, C, H, W] -> [B, C, F, H, W]
+#         cond_latents = pipe.vae.encode(image_or_video).latent_dist.sample()
+#         cond_latents = cond_latents * pipe.vae.config.scaling_factor
+#         cond_latents = cond_latents.to(dtype=pipe.dtype)
+#         assert not torch.any(torch.isnan(cond_latents))
+    
+#     # Generate video
+#     video = call_pipe(
+#         pipe,
+#         prompt=prompt,
+#         num_frames=num_frames,
+#         num_inference_steps=num_inference_steps,
+#         image_latents=cond_latents,
+#         width=width,
+#         height=height,
+#         guidance_scale=guidance_scale,
+#         generator=torch.Generator(device="cuda").manual_seed(0),
+#     ).frames[0]
+    
+#     # Export to video
+#     video_path = "output.mp4"
+#     export_to_video(video, video_path, fps=24)
+    
+#     with open(video_path, "rb") as video_file:
+#         video_bytes = video_file.read()
+    
+#     return video_bytes
 
 @torch.inference_mode()
 def call_pipe(