update

app.py

@@ -202,7 +202,9 @@ pipeline = XoraVideoPipeline(
     tokenizer=tokenizer,
     scheduler=scheduler,
     vae=vae,
-).to(device)
+)
+#).to(device)
+pipeline.enable_model_cpu_offload()
 
 
 def generate_video_from_text(

requirements.txt

@@ -6,7 +6,7 @@ sentencepiece>=0.1.96
 accelerate
 einops
 matplotlib
-opencv-python
+opencv-python==4.5.5
 beautifulsoup4
 ftfy
 gradio

xora/models/autoencoders/vae_encode.py

@@ -127,6 +127,7 @@ def _run_decoder(
         *_, fl, hl, wl = latents.shape
         temporal_scale, spatial_scale, _ = get_vae_size_scale_factor(vae)
         latents = latents.to(vae.dtype)
+        vae = vae.to(latents.device, vae.dtype)
         image = vae.decode(
             un_normalize_latents(latents, vae, vae_per_channel_normalize),
             return_dict=False,