This AI Deepfake Went Viral – This is How to Make it

AI deepfakes are rapidly evolving, and recently, a viral deepfake video by original Max took the internet by storm.

Key Takeaways:

Learn how to create viral AI deepfakes using Stable Diffusion.
Discover the specific models, LoRAs, and embeddings needed for realistic results.
Understand how to set up ControlNet for seamless consistency between frames.
Follow a step-by-step guide on transferring data from Img2Img to Deforum.
Get insights on improving your workflow with free resources like settings files.

This tutorial will show you step-by-step how to replicate that deepfake video using Stable Diffusion and various models, LoRAs, and embeddings. By following along, you’ll learn how to create stunning AI videos that feature seamless character transitions, just like the viral deepfakes you’ve seen.

The Rise of AI Deepfakes

AI deepfakes have become increasingly popular, with creators experimenting with advanced techniques to blend different characters into seamless videos. In this viral video, characters transition smoothly from Cristiano Ronaldo to Arnold Schwarzenegger, Brad Pitt, and even George Clooney. In this post, I’ll show you how to replicate this effect using Stable Diffusion. You’ll be creating deepfakes that rival the best in no time!

I’ll also share some tips and tools that were generously shared by Andre from Unreal Unit, who helped make this tutorial possible.

Setting Up Stable Diffusion for AI Deepfake Creation

Before diving into Stable Diffusion, let’s look at the models, LoRAs, and embeddings you’ll need for this deepfake project. These assets will ensure your characters look accurate and consistent throughout the video. You can find the links to download them at the end of this post.

Deliberate V2 Model: The foundation for generating high-quality images.
LoRAs: Lionel Messi, Brad Pitt, Ryan Reynolds, and Tom Hanks.
Embeddings: Bad Hand V4, Easy Negative, Very Bad Image Negative.

Once you have these downloaded, you can install them by placing the files in the correct folders. If you’re new to model and embedding installation, check out the video where I explain how to install them at [4:52] and [7:28].

Creating Prompts for Each Character in Img2Img

Now that you have everything set up, it’s time to create prompts for each character in Stable Diffusion’s Img2Img tab. You’ll generate the base images for each character before transitioning them in Deforum.

Step-by-Step Instructions:

Drag your image into the Img2Img tab.
Set the model to Deliberate.
Configure the following settings for each character:
- Sampling method: Euler A
- Sampling steps: 20
- Width: 576
- Height: 1024
- CFG Scale: 4
- Denoising strength: 1

ControlNet Settings:

ControlNet 1:
- Enable Pixel Perfect.
- Set Control Type to Tile.
- Preprocessor: None.
- Model: Control Tile.
- Control Weight: 1.5.
- Control Mode: ControlNet is More Important.
ControlNet 2:
- Enable.
- Control Type: All.
- Preprocessor: None.
- Model: Temporal Net.
- Control Weight: 0.8.
- Control Mode: ControlNet is More Important.

Ensure the Noise Multiplier is set to 0.05 for each character. If you have trouble setting it up, watch the troubleshooting video at [3:46].

Assigning Prompts to Each Character

Frames 0-138: Lionel Messi (using Messi LoRA).
Frames 138-284: Brad Pitt (using Brad Pitt LoRA).
Frames 284-394: Ryan Reynolds (using Ryan Reynolds LoRA).
Frames 394-494: Tom Hanks (using Tom Hanks LoRA).

Make sure your negative prompt is consistent across all characters and that you don’t forget to remove any extra commas at the end of your prompt list, as it could cause errors during rendering.

Transferring Img2Img Data into Deforum

Now that you’ve set up the individual character prompts, it’s time to move the data into Deforum for video creation. This will allow you to transition between characters seamlessly.

Load Base Deforum Settings: You can download my free settings file from Ko-Fi. Once you have it, load the settings by copying the path and pasting it into Deforum.
Run Tab:
- Sampler: Euler A.
- Steps: 20.
- Width: 576.
- Height: 1024.
Keyframes Tab:
- Set CFG Schedule to 4.

In the Prompts Tab, assign the prompts for each character, following the sequence from Img2Img. You’ll want to ensure smooth transitions between characters by tweaking the start and end times for each prompt.

Finalizing the Deepfake in Deforum

Once all the prompts and settings are in place, you’re ready to generate the video. Make sure your output settings match your base video’s resolution and framerate.

After generating the video, you can enhance it further with upscaling and post-processing.

Tips for Post-Processing:

Use frame interpolation to improve the flow between frames.
Add a dirt removal node and deflicker node in DaVinci Resolve.

Check out the video at [15:58] for a walkthrough of the post-production process.

Conclusion

And there you have it! By following this guide, you’ll be able to create stunning AI deepfake videos using Stable Diffusion and Deforum, just like the viral videos from creators like original Max and Unreal Unit. Don’t forget to download the free settings files I’ve provided to speed up your workflow, and be sure to experiment with different ControlNet settings to achieve the best results.