VAE encode fixes

hyvideo/diffusion/pipelines/pipeline_hunyuan_video.py

@@ -512,7 +512,7 @@ class HunyuanVideoPipeline(DiffusionPipeline):
         self._num_timesteps = len(timesteps)
 
         
-        logger.info(f"Sampling {video_length} frames in {latents.shape[2]} latents at {width}x{height} with {num_inference_steps} inference steps")
+        logger.info(f"Sampling {video_length} frames in {latents.shape[2]} latents at {width}x{height} with {len(timesteps)} inference steps")
         comfy_pbar = ProgressBar(len(timesteps))
         with self.progress_bar(total=len(timesteps)) as progress_bar:
             for i, t in enumerate(timesteps):

hyvideo/vae/autoencoder_kl_causal_3d.py

@@ -486,9 +486,9 @@ class AutoencoderKLCausal3D(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
 
         B, C, T, H, W = x.shape
         overlap_size = int(self.tile_sample_min_tsize *
-                           (1 - self.tile_overlap_factor))
+                           (1 - self.t_tile_overlap_factor))
         blend_extent = int(self.tile_latent_min_tsize *
-                           self.tile_overlap_factor)
+                           self.t_tile_overlap_factor)
         t_limit = self.tile_latent_min_tsize - blend_extent
 
         # Split the video into tiles and encode them separately.

nodes.py

@@ -893,7 +893,7 @@ class HyVideoDecode:
                 vae.t_tile_overlap_factor = 0.25
         else:
             #defaults
-            vae.tile_latent_min_tsize = 64
+            vae.tile_latent_min_tsize = 16
             vae.tile_sample_min_size = 256
             vae.tile_latent_min_size = 32
         
@@ -976,7 +976,7 @@ class HyVideoEncode:
                 vae.t_tile_overlap_factor = 0.25
         else:
             #defaults
-            vae.tile_latent_min_tsize = 64
+            vae.tile_latent_min_tsize = 16
             vae.tile_sample_min_size = 256
             vae.tile_latent_min_size = 32