diff --git a/diffsynth/pipelines/sd_video.py b/diffsynth/pipelines/sd_video.py
index eb812cf..8a577b5 100644
--- a/diffsynth/pipelines/sd_video.py
+++ b/diffsynth/pipelines/sd_video.py
@@ -257,10 +257,10 @@ class SDVideoPipeline(SDImagePipeline):
                 progress_bar_st.progress(progress_id / len(self.scheduler.timesteps))
         
         # Decode image
-        image = self.decode_video(latents, **tiler_kwargs)
+        output_frames = self.decode_video(latents, **tiler_kwargs)
 
         # Post-process
         if smoother is not None and (num_inference_steps in smoother_progress_ids or -1 in smoother_progress_ids):
             output_frames = smoother(output_frames, original_frames=input_frames)
 
-        return image
+        return output_frames
diff --git a/diffsynth/pipelines/sdxl_video.py b/diffsynth/pipelines/sdxl_video.py
index d918ae7..faa8bff 100644
--- a/diffsynth/pipelines/sdxl_video.py
+++ b/diffsynth/pipelines/sdxl_video.py
@@ -214,10 +214,10 @@ class SDXLVideoPipeline(SDXLImagePipeline):
                 progress_bar_st.progress(progress_id / len(self.scheduler.timesteps))
         
         # Decode image
-        image = self.decode_video(latents, **tiler_kwargs)
+        output_frames = self.decode_video(latents, **tiler_kwargs)
 
         # Post-process
         if smoother is not None and (num_inference_steps in smoother_progress_ids or -1 in smoother_progress_ids):
             output_frames = smoother(output_frames, original_frames=input_frames)
 
-        return image
+        return output_frames
diff --git a/examples/ExVideo/ExVideo_svd_train.py b/examples/ExVideo/ExVideo_svd_train.py
index 4342315..6a6221e 100644
--- a/examples/ExVideo/ExVideo_svd_train.py
+++ b/examples/ExVideo/ExVideo_svd_train.py
@@ -4,7 +4,7 @@ import numpy as np
 from einops import rearrange, repeat
 import lightning as pl
 from diffsynth import ModelManager, SVDImageEncoder, SVDUNet, SVDVAEEncoder, ContinuousODEScheduler, load_state_dict
-from diffsynth.pipelines.stable_video_diffusion import SVDCLIPImageProcessor
+from diffsynth.pipelines.svd_video import SVDCLIPImageProcessor
 from diffsynth.models.svd_unet import TemporalAttentionBlock
 
 
@@ -131,14 +131,14 @@ class LightningModel(pl.LightningModule):
         self.image_encoder.requires_grad_(False)
 
         self.unet = SVDUNet(add_positional_conv=add_positional_conv).to(dtype=torch.float16, device=self.device)
-        self.unet.load_state_dict(SVDUNet.state_dict_converter().from_civitai(state_dict), strict=False)
+        self.unet.load_state_dict(SVDUNet.state_dict_converter().from_civitai(state_dict, add_positional_conv=add_positional_conv), strict=False)
         self.unet.train()
         self.unet.requires_grad_(False)
         for block in self.unet.blocks:
             if isinstance(block, TemporalAttentionBlock):
                 block.requires_grad_(True)
 
-        self.vae_encoder = SVDVAEEncoder.to(dtype=torch.float16, device=self.device)
+        self.vae_encoder = SVDVAEEncoder().to(dtype=torch.float16, device=self.device)
         self.vae_encoder.load_state_dict(SVDVAEEncoder.state_dict_converter().from_civitai(state_dict))
         self.vae_encoder.eval()
         self.vae_encoder.requires_grad_(False)
diff --git a/examples/image_synthesis/sd3_text_to_image_textual_inversion.py b/examples/image_synthesis/sd3_text_to_image_textual_inversion.py
deleted file mode 100644
index 1cf5256..0000000
--- a/examples/image_synthesis/sd3_text_to_image_textual_inversion.py
+++ /dev/null
@@ -1,32 +0,0 @@
-from diffsynth import ModelManager, SD3ImagePipeline, download_models, load_state_dict
-import torch
-
-
-# Download models (automatically)
-# `models/stable_diffusion_3/sd3_medium_incl_clips.safetensors`: [link](https://huggingface.co/stabilityai/stable-diffusion-3-medium/resolve/main/sd3_medium_incl_clips.safetensors)
-# `models/textual_inversion/verybadimagenegative_v1.3.pt`: [link](https://civitai.com/api/download/models/25820?type=Model&format=PickleTensor&size=full&fp=fp16)
-download_models(["StableDiffusion3_without_T5", "TextualInversion_VeryBadImageNegative_v1.3"])
-model_manager = ModelManager(torch_dtype=torch.float16, device="cuda")
-model_manager.load_textual_inversions("models/textual_inversion")
-model_manager.load_models(["models/stable_diffusion_3/sd3_medium_incl_clips.safetensors"])
-pipe = SD3ImagePipeline.from_model_manager(model_manager)
-
-
-for seed in range(4):
-    torch.manual_seed(seed)
-    image = pipe(
-        prompt="a girl, highly detailed, absurd res, perfect image",
-        negative_prompt="verybadimagenegative_v1.3",
-        cfg_scale=4.5,
-        num_inference_steps=50, width=1024, height=1024,
-    )
-    image.save(f"image_with_textual_inversion_{seed}.jpg")
-
-    torch.manual_seed(seed)
-    image = pipe(
-        prompt="a girl, highly detailed, absurd res, perfect image",
-        negative_prompt="",
-        cfg_scale=4.5,
-        num_inference_steps=50, width=1024, height=1024,
-    )
-    image.save(f"image_without_textual_inversion_{seed}.jpg")