I just finished creating a new Tom and Jerry scene and the result is amazing. This clip is over 40 seconds long and the character looks exactly right in every single frame. Even when I reach the 40 second mark the face and the small details stay consistent. It feels like watching a real movie because the video is long but the character does not change or morph at all. I managed to generate this entire sequence with just one click and I want to show you exactly how you can do it too.
File You Need To Download
Before we start I have to talk about the files you need to get this workflow running. I am using the Smooth Mix Wan 2.2 model which is a special fine tuned version of the standard Wan 2.2. Instead of using the standard Wan 2.2 model you should download these files:
- Smooth Mix Wan 2.2 I2V High Noise: https://civitai.com/models/1995784?modelVersionId=2260110
- Smooth Mix Wan 2.2 I2V Low Noise: https://civitai.com/models/1995784?modelVersionId=2259006
- Rank 64 LightX2V LoRA (Latest Dec 17th version) lightx2v/Wan2.2-Distill-Loras at main
- SVI V2 Pro I2V LoRA https://huggingface.co/vita-video-gen/svi-model/tree/main/version-2.0
- Wan 2.2 Text Encoder and VAE
These specific models are much better at handling motion compared to the original
. The best part is that you can generate great results in just 4 to 8 steps even without using any extra lora files.
Why You Must Delete Inspector Blend
I want to give you a very important tip that most of the community misses. When you open an older SVI Pro workflow you might see a node called Free Long Inspector Blend. I know many people in the community use this because they think it helps keep things consistent for long videos but it actually causes a big problem. This node restricts the natural motion of the character and makes the entire video look stiff or static. It basically fights against the smooth motion of the Wan model. I also tested the Free Long Enforcer node and I found that it just kills the movement and adds no value. I get a much cleaner and smoother result without them so I recommend you delete these nodes immediately.
Setting Up For Success
Now we are ready to generate a scene. For my test I uploaded an image of a woman sitting on a bed and I used a resolution of 720 by 1280. I am running this on a 5090 GPU and it takes about 1 or 2 minutes to finish. If your GPU doesn’t have much VRAM you can bring the resolution down to 480 by 832. There is also a GGUF version of Smooth Mix available for download. If you have low RAM just use that GGUF version and it will run much better on your system.+4
One setting you have to check is the Shift value inside the Model Sampling section. The default is usually 5 but if you are making a long video and notice the colors starting to drift after 20 seconds you have to change it. I increase the Shift value from 5 to 12 and that completely fixes the color drift while keeping the motion perfect.
Wan 2.2 SVI2 Pro Workflow Guide for Long AI Videos
I just finished creating a new Tom and Jerry scene and the result is amazing. This clip is over 40 seconds long and the character looks exactly right in every single frame. Even when I reach the 40 second mark the face and the small details stay consistent. It feels like watching a real movie because the video is long but the character does not change or morph at all. I managed to generate this entire sequence with just one click and I want to show you exactly how you can do it too.
File You Need To Download
Before we start I have to talk about the files you need to get this workflow running. I am using the Smooth Mix Wan 2.2 model which is a special fine tuned version of the standard Wan 2.2. Instead of using the standard Wan 2.2 model you should download these files:
- Smooth Mix Wan 2.2 I2V High Noise
- Smooth Mix Wan 2.2 I2V Low Noise
- Rank 64 LightX2V LoRA (Latest Dec 17th version)
- SVI V2 Pro I2V LoRA
- Wan 2.2 Text Encoder and VAE
These specific models are much better at handling motion compared to the original
. The best part is that you can generate great results in just 4 to 8 steps even without using any extra lora files.
Why You Must Delete Inspector Blend
I want to give you a very important tip that most of the community misses. When you open an older SVI Pro workflow you might see a node called Free Long Inspector Blend. I know many people in the community use this because they think it helps keep things consistent for long videos but it actually causes a big problem. This node restricts the natural motion of the character and makes the entire video look stiff or static. It basically fights against the smooth motion of the Wan model. I also tested the Free Long Enforcer node and I found that it just kills the movement and adds no value. I get a much cleaner and smoother result without them so I recommend you delete these nodes immediately.
Setting Up For Success
Now we are ready to generate a scene. For my test I uploaded an image of a woman sitting on a bed and I used a resolution of 720 by 1280. I am running this on a 5090 GPU and it takes about 1 or 2 minutes to finish. If your GPU doesn’t have much VRAM you can bring the resolution down to 480 by 832. There is also a GGUF version of Smooth Mix available for download. If you have low RAM just use that GGUF version and it will run much better on your system.+4
One setting you have to check is the Shift value inside the Model Sampling section. The default is usually 5 but if you are making a long video and notice the colors starting to drift after 20 seconds you have to change it. I increase the Shift value from 5 to 12 and that completely fixes the color drift while keeping the motion perfect.
Using Negative Prompts Correctly
I ran into a small problem when I wanted her to open the white door. The model kept trying to open the brown wardrobe again because it thought that was the only door in the room. Instead of fighting with the positive prompt I just added brown cabinet to the Negative Prompt. That simple trick worked perfectly and the model ignored the cabinet to follow my command. If you are struggling with a specific object just put it in the negative section and it will save you a lot of time.
Finishing The Long Clip
As the camera follows her into the living room the lighting stays warm and consistent. I even made her go outside the house to see how the model handles natural light. I add the prompt for bright light to flood in when she opens the door and the transition looks great. We reached 461 frames which is about 26 seconds of video with zero color distortion.+2
At the very end of the 40 second clip I wanted to check if her face was still the same. I stopped the camera and had her turn 180 degrees to look directly at the lens. Her face matched the original reference image perfectly. It matches exactly like the original picture. If you want to keep extending your video just copy all the groups and connect the Extended Image nodes. You can keep doing this to make your video as long as you want.