minor

app.py

@@ -76,16 +76,17 @@ def generate(prompt, num_frames, image, model_name_stage1, model_name_stage2, se
 
     n_autoreg_gen = (num_frames-8)//8
 
-    inference_generator = torch.Generator(device="cuda").manual_seed(seed)
-
     if model_name_stage1 == "ModelScopeT2V (text to video)":
+        inference_generator = torch.Generator(device=ms_model.device).manual_seed(seed)
         short_video = ms_short_gen(prompt, ms_model, inference_generator, t, device)
     elif model_name_stage1 == "AnimateDiff (text to video)":
+        inference_generator = torch.Generator(device=ad_model.device).manual_seed(seed)
         short_video = ad_short_gen(prompt, ad_model, inference_generator, t, device)
     elif model_name_stage1 == "SVD (image to video)":
         # For cached examples
         if isinstance(image, dict):
             image = image["path"]
+        inference_generator = torch.Generator(device=svd_model.device).manual_seed(seed)
         short_video = svd_short_gen(image, prompt, svd_model, sdxl_model, inference_generator, t, device)
 
     stream_long_gen(prompt, short_video, n_autoreg_gen, seed, t, image_guidance, name, stream_cli, stream_model)

t2v_enhanced/model_init.py

@@ -119,6 +119,7 @@ def init_v2v_model(cfg, device):
 
     pipe_enhance.model.autoencoder = pipe_enhance.model.autoencoder.to(device)
     pipe_enhance.model.generator = pipe_enhance.model.generator.to(device)
+    pipe_enhance.model.generator = pipe_enhance.model.generator.half()
     pipe_enhance.model.negative_y = pipe_enhance.model.negative_y.to(device)
     pipe_enhance.model.cfg.max_frames = 10000
     return pipe_enhance