AI art filters have become incredibly popular on TikTok in recent years. These filters use artificial intelligence to apply different artistic effects and transformations to videos and images. Some of the most popular AI art filters on TikTok include Cartoonify,Vincent Van Gogh filter, Anime filter, and Sketch filter. These filters use neural networks to analyze the visual contents of an image or video frame by frame and apply an artistic style or effect. The AI has been trained on vast datasets of artworks to understand different artistic techniques like brush strokes, colors, textures and more. This allows the AI filter to transform regular videos into unique works of art. The AI art filters on TikTok have become a sensation, allowing users to easily give their content a stunning painted or cartoon look. Their popularity stems from the fact that they make video creation fun and engaging while opening up new creative possibilities.
Source: https://www.techhivey.com/2023/07/reverse-ai-art-filter-tiktok.html
How AI Art Filters Work
TikTok’s AI art filters utilize a form of artificial intelligence called generative adversarial networks, or GANs. GANs are made up of two neural networks – a generator and a discriminator. The generator creates new images while the discriminator evaluates them, sending feedback to the generator on how to improve. This adversarial training process allows the GAN to become extremely adept at generating realistic images.
Specifically, TikTok likely uses a type of GAN known as StyleGAN, pioneered by Nvidia researchers. StyleGAN generates highly realistic and diverse images by separating aspects of image synthesis into different levels of abstraction. For example, it has high-level controls for pose and identity, mid-level controls for lighting and color, and low-level controls for specific facial features.
By training StyleGAN on a dataset of human faces, TikTok’s engineers can produce an AI model capable of generating an endless variety of realistic human portraits. This model powers the AI art filters that transform users’ selfies in the app. According to TikTok’s guide, creators can tap into these generative AI models to easily create new effects.
Why Reverse the Filters?
There are several reasons why people may want to reverse the AI art filters on TikTok to remove the effects and see the original image:
To get back the original photo if the edited version is too distorted or unrecognizable. The filters can sometimes alter images drastically, making subjects unrecognizable. Reversing the filter returns the image to its original state (Source).
For fun and entertainment. Trying to guess what the original photo looked like based on the heavily edited version, then checking by reversing the filter, has become a popular trend (Source).
To aid mental health and self-exploration. The dreamlike distortions created can be used as a tool for examining emotions, thoughts, and self-perception in a new light (Source). Reversing the filter returns the familiar after this self-reflection.
To critique the technology.Examining differences between the original and edited versions allows consideration of how the AI algorithm works, its capabilities and limitations.
Built-in Option to Reverse?
As of now, TikTok does not have a built-in option to reverse the AI art filters. The AI art filters use neural networks to transform images in artistic styles, which is a complex process. There is no simple button to press to revert the images back to their original state.
Some users have experimented with applying filters on top of the AI art images to try to reverse the effect, but this does not reliably restore the original image. The AI transformations result in too much alteration of colors, textures, and details to be easily reversed.
While TikTok could theoretically develop a feature to reverse the AI art filters by training another neural network model, this would require significant engineering work. Given the complexity, TikTok has not yet invested resources into building a reversal feature.
Users wishing to restore their original images after applying AI art filters will need to find alternative solutions for now. TikTok may eventually provide official tools, but reversing AI art requires more than a simple filter.
Third Party Apps
While TikTok doesn’t have a built-in option to reverse AI art filters, there are some third party apps that aim to provide this functionality. One example is Anthropic’s Stable Diffusion Playground (https://www.anthropic.com), which allows users to upload a TikTok video or image and attempt to reverse the filter to recreate the original. The app uses a deep learning model called Stable Diffusion to analyze the filtered image and generate an unfiltered version. However, the results can be mixed and it doesn’t always perfectly recreate the original.
Another option is Runway (https://runwayml.com), an AI assistant that offers various image generation capabilities. Users can upload a filtered TikTok image and leverage Runway’s GAN inversion feature to attempt reversing the filter. This uses a generative adversarial network to reconstruct the original image from the stylized version. Like Stable Diffusion Playground, it doesn’t guarantee perfect results. There are also paid services like Img2Img (https://img2img.com) that provide more advanced filter reversal capabilities, but free apps have limited effectiveness.
Image Processing
Image processing techniques like convolution can be used to reverse engineer and remove filter effects applied to images. Convolution is an algorithm commonly used in image processing that applies a kernel or filter to an input image to produce an output image. By analyzing the filtered image and approximating the type of kernel or filter applied, it may be possible to reconstruct the original, unfiltered image.
One technique is to apply an approximated inverse filter to attempt to cancel out the original filter. Wang et al. (2023) demonstrated this on simulated noisy and blur filters, showing it was possible to reconstruct sharper and clearer images [1]. However, perfectly inverting filters is challenging, especially when details of the original filter are unknown.
Alternatively, iterative optimization methods can be used to gradually refine the image by comparing it to the filtered output, until a close approximation of the original is obtained. Belyaev et al. (2021) proposed two such iterative techniques that were able to significantly reduce strong filtering effects like pixelation and jpg compression artifacts [2].
While showing promising results, these image processing techniques have limitations. It is difficult to perfectly reconstruct images from heavily filtered outputs, especially when some irreversible loss of information has occurred. The original unfiltered image may be approximated but some artifacts are likely to remain.
Frame Interpolation
One potential method to reverse the AI art filter effect on videos is through frame interpolation. Frame interpolation is a technique in image and video processing where new frames are generated to be inserted between existing ones. This allows for increasing the frame rate and making the video appear smoother.1
By running the TikTok video through advanced frame interpolation algorithms, it may be possible to reconstruct frames that existed before the filter was applied. Essentially, the algorithm would analyze the filtered video, estimate the original frames, and produce an interpolation that recreates the video prior to filtering.
Frame interpolation AIs like DAIN and BMBC can increase frame rates up to 120 fps and provide smooth slow motion effects. Researchers have been investigating reversing this process to reconstruct static images from video. While challenging, initial results suggest AI-based frame interpolation could help remove or isolate filter effects in some cases.2
However, there are limits to this technique. If the filter makes extensive changes to each frame, interpolation may be unable to reliably estimate the originals. More research is needed into inversion techniques tailored to common filter effects seen on TikTok and social media.
GAN Inversion
One method for reversing AI art filters is called GAN inversion. GAN stands for Generative Adversarial Network. GANs are a type of neural network architecture that can generate new images that appear authentic to human viewers.
GAN inversion involves using the GAN model in reverse – instead of generating a new image, the goal is to reconstruct the original image from the filtered version. This is done by optimizing and iterating on the latent vector input to the GAN until the output matches the target filtered image as closely as possible.[1]
By slowly adjusting the latent vector, the GAN can “invert” the image back towards the original. This takes advantage of the fact that GANs encode images into a compressed latent space representation. So the inverse mapping from output to latent vector can approximate reconstructing the original input image.[1]
However, GAN inversion has limitations – the reconstruction may not perfectly match the original image. The compression of the latent space loses some image information. But GAN inversion can get reasonably close to reconstructing the original from the filtered image.
Limitations
While reversing AI art filters is possible, there are some challenges and limitations to be aware of:
One major limitation is that full reversal may not always be achievable. According to research from https://siglets.com/reversing-ai-art-filters/, current techniques can generally recover the overall structure and content of the original image, but some finer details may be lost or distorted in the process.
Another challenge is that different AI filters use different techniques, so there is no universal reversal method that works perfectly for all filters. The reversal process needs to be tailored to the specific filter that was originally applied.
In some cases, the original image may be altered or degraded too much by the filter, making high fidelity reversal difficult. This is especially true for filters that apply heavy artistic stylization or distortions.
There are also limitations around proprietary AI systems where the details of the filters are not publicly known. Without knowledge of the system’s inner workings, it can be very difficult to inverse its effects.
While active research is making progress in this area, reversing AI art filters remains an imperfect process. The quality and completeness of the reversal depends heavily on the technique used and the nature of the original filter.
Conclusion
In summary, the AI art filters on TikTok alter images and videos in an irreversible way through complex neural networks. While there are some emerging techniques like GAN inversion that may be able to reverse engineered images in the future, currently there is no built-in option in TikTok to reverse filters and completely recover the original.
Third party apps can somewhat simulate reversing the effects through interpolation and processing, but they cannot perfectly reconstruct the original. The AI filters fundamentally change and lose pixel data from the source. While reversing the effects may be possible by training specialty AI models on large datasets in the future, for now once an image or video has passed through TikTok’s filters there is no way to get the exact pre-filtered version back.
In conclusion, while some promising research is being done, at this time it is not feasible to fully reverse engineered TikTok’s AI art filters back to the original input image or video.
