Improve video quality by removing unwanted camera shakes and jitters due to hand shaking and unintentional camera panning. This model stabilizes footage by moving and rotating the original frames to make subjects visibly stable from frame to frame.

Input and output files will not match pixel to pixel when using this model.

Increasing strength will result in a more stabilized output at the cost of increased artifacts (Full Frame) or a more cropped output (Auto-Crop).

Full-frame mode fills the blank spaces at the edges of the video using frame interpolation and deep learning.

This method preserves the original resolution but may result in blurriness on edges in some cases. 

Auto Crop mode will crop out blank spaces on the edges of videos, resulting in an image that is always sharp but may have a reduced resolution and frame size. In the example above, the striped portion is what will be cropped.

Larger portions of the original frame may be removed depending on the amount of camera shake being stabilized.

In the example above, the border of filmed data, is the visual difference between Full Frame and Auto Crop. Auto Crop completely removes a safe percentage of data that may contain artifacts from the stabilized motion. As noted, the resulting crop is based upon the amount of camera shake that is present in the original footage.

Reduces wobbly distortion caused when the camera moves too fast. 

Here is an example of the same two images lined up and overlayed. There is a clear distortion with the angle and stretched appearance of the bird.

This toggle corrects parallax distortion and tilting motion by running Chronos Fast for each interpolated pass.

Additional passes will significantly increase processing time.

Have fun stabilizing that shaky cam with these options!

Stabilization

Reduces the blurring caused by objects or the camera in quick motion. This tool is great on a multitude of footage types and situations. Motion Deblur was built with a wide range of circumstances in mind.

The clip above shows the result of a bumpy auto rickshaw ride and how the Motion Deblur model improves the footage when the camera is moving quickly.

For the penguin clip above, the camera is following the animal in place as the subject is swimming quite quickly in the footage. Details are preserved, even with the penguin’s speed.

Themis is the AI Model built for Motion Deblur. It increases the perceived sharpness and detail of videos by removing blur instead of just increasing edge contrast. Themis reduces blur caused by panning, rotation or zoom.

Motion Deblur is accessible just below “Stabilization” from the Controls Sidebar.

Click on the slider button to reveal the dropdown menu.

Once Motion Deblur is turned “on” from the slider, the effect will be applied instantly.

Motion Deblur

Ever wished that you could refine time and space? Well here’s a tool that can get your footage there!

Frame Interpolation, is the process of synthesizing new frames from a given set of frames to create more fluid movement or to slow down the speed of a video. 

Frame Rates are essentially the number of frames that are displayed consecutively per a calculated length of time. The widely accepted identifier is “Frames Per Second” (FPS). The amount of frames that are shown per second plays a large role on how we perceive a video in not just the speed, but in creating a more comfortable experience or emoting the sense of a nostalgic era of film to name a few.

16mm (Silent): Generally shot around 16-18 FPS

Cinematic Action Sequences: Generally shot around 60-120 FPS

Frame rates can be changed by selecting the caret and navigated down the selection of an existing frame rate.

If you want to make modify the existing frame rates and make a custom one, you can simply type in a desired framerate.

Please note: Frame rate limitation can be determined by the codec your are using.

Post or Digital Slow Motion is the process of interpreting the original frames and duration of a video to algorithmically exceed the current length by a determined value, from the user, in order to produce a longer and slower video iteration of the original input. In other words, you can digitally slow down your original footage to create a result that was not recorded in reality.

Choose a slow motion factor, or to keep the speed the same, select 

 will replace duplicate frames with interpolated frames.

 can be raised to decrease the threshold for removing duplicate frames. A higher setting may remove frames that are not duplicates but look very similar.

Please note: Frame interpolation is only applied to the video. The audio will not be modified by this enhancement and will be removed from the video.

Specializes in non-linear motion and slightly blurry inputs.

Generates 8 interpolated frames per run, and then creates the selected output framerate from this result.

 frame count multiple. For example, converting from 30 → 60fps at 200% slow motion would require exactly 4x the number of frames.

Apollo worked wonderfully for this (slow motion 4x). This is a handheld, but nearly locked shot.

The image above is only a visual reference - not in-app

On the original footage, the water is rushing very quickly. Using this filter produced a wonderful result of the churning white water. The additional movement of the wind pushing through the leaves and plants resulted nicely, but the water is the true star of this shot.

Typically interpolates runs faster than Apollo, however, may produce a lesser quality output than Apollo when working with large amount of motion.

Recommended for use for any non-double conversions. For example, 24fps -> 29.97 or 60.

Though the quality is slightly reduced compared to the standard Apollo, for a busy shot like this, the difference is not crucial to the intended effect of the filter. 

It may be worth testing Apollo Fast when rendering long sequences. The result overall is still quite good, and could save significant amounts of rendering time.

General frame interpolation model suitable for FPS conversion or slow-mo. Processing speed decreases linearly with the number of generated frames.

If render speed is a factor for slow motion beyond 8x this is a good option if the Aion model is too time and system-intensive. 

Gives good results for 4x slow motion or less.

Being that Chronos is one of the app's earlier models, it is also one of the more general models that handles footage relatively well. For this shot, the boats are rocking in the water, there are reflections under the boats, and there are thin ropes to each mast. The results were pretty good even with something as drastic as slow motion 10x.

Specializes in fast motion with larger changes between frames.

Can process faster than Chronos, but can be very similar at times as well. Worth doing a small test for render time savings.

Using the same shot as Chronos, there are a few visual differences on how these work. The area of interest will be this small section of the water.

At a very close comparison the differences are quite noticeable.

In the “red” rectangular portion, Chronos Fast is treating the footage like scanlines or fields to quickly get the result. Chronos overall looks better in this area. However there are spots in the Chronos were it also has these very thin vertical scanlines. It’s worth noting that we are very zoomed into the footage to be able to see these differences.

High-precision interpolation for large media

More accurate than Apollo and Chronos, but more system-intensive.

Useful when working with large amounts of motion across the entire frame of an input video.

Here we have a simple panning shot to the left, handheld, but nearly locked off. There are several levels of parallax and Aion was able to handle this very well.

The original footage of the city moved abruptly and there were only a few moments of really stellar footage. With the Aion model, the few moments were expanded by 8 times. This is a very wide shot of New York City behind a glass wall. Even with the glass in front, the result was still great. All the tiny details were preserved.

Here we are looking at natively shot stereoscopic 180' footage for VR. Aion was able run the footage at 12 times and since it is paramount that both eyes are addressed in a similar way for reduced stereo artifacts, this model was able to provide the slowdown in equal quality in both eyes.

Frame Interpolation

Enhancement filters are the most customizable feature of Topaz Video AI. This section contains many controls to refine and improve your footage. 

This will be your guide through the Enhancement filter's settings, with our recommendations for different types of input files. 

Please note, every piece of footage is unique and we encourage experimentation with different models and parameters as part of the creative process. So have fun and start experimenting!

You can find the Enhancement menu on the screen right side in the Controls Sidebar.

To open the Enhancement controls, click the activation toggle.

The Output Resolution section allows you to select the final dimensions of your export. Click on the dropdown menu to reveal all the options available, including a custom selection at the bottom.

In the Video Type section, you can set the Enhancement filter to match the type of footage you are processing. Changing this setting will list AI models that can be utilized with the selected Video Type.

 Enhance and upscale without deinterlacing. Most modern digital videos are progressive.

 Deinterlace, enhance, and upscale footage with a combined model.

 These filters are trained for interlaced videos that were previously transcoded to progressive without deinterlacing.

, there are now recommended models based on the input resolution of your video. There is also a description below the visual carousel for selecting AI models.

To switch to a more compact list view, click the three horizontal lines to the right of the panel. 

Each AI model was designed with certain tasks in mind. That being said, it is worth experimenting with each of the models to see what each option is optimized for.

Most of Video AI's models have the following three modes:

 Determines the optimal settings based on the input video's characteristics.

 This will determine the optimal settings and add/reduce the strength of each setting based on slider placement.

 to allow the app to generate automatic settings as a starting guide for customization.

 Adds digital noise to counter texture smoothing as a result of noise reduction.

 This slider controls how much original source detail will be mixed back into the enhanced footage.

 can assist in some out of focus video content. This tool downscales, then upscales to the video’s resolution. By default, this is set to 

 downscales at 50% and upscales at two times the scale.

 downscales at 25% and upscales four times the scale.

Complex enhancements can sometimes benefit from stacking two Enhancement filters together. Are you ready for this advanced function? 

An interlaced file uses an early form of video compression that was used to make videos look smoother and reduce bandwidth requirements on early CRT televisions. Interlaced video is commonly produced by older cameras, visually the frames will show horizontal rows throughout called 

. Every other field of data captures one of two different moments of time. Interlaced file types are often represented with the letter “i” next to the format. Examples of this would be 480i, 720i, or 1080i.

The Dione models are tuned to deinterlace interlaced files and upscale using a single AI model. Every interlaced model allows you to select the Field order and add noise.

Add Noise simply allows the user to add digital noise to counter some strong noise reduction and texture smoothing results.

Dione DV is trained for digital video footage like DV, DVCPRO, DVCAM, 720i, and 1080i. 

Dione TV is trained for analog videos from older TVs, camcorders, DVDs, and SD footage (480i, 576i, etc.). This model has additional dehaloing. This helps with common halo issues in these older formats and slightly lifts shadows for more clarity in high contrast videos.

Dione TV variant that specializes in removing halo and fringing. Similar to Dione TV, this model helps with halo and high contrast issues common in these older formats.

Though these three Dione models have specific use cases in mind, it's worth the time to run interlaced video through all three models. Here is a visual example of what to expect from each variant.

Sometimes having a little extra dehalo on a miniDV clip renders a nice look.

It can give a subtle lift and softness through the reduction of contrast and added slight haze. It creatively gives a little atmosphere to the shot. For the shot above, it has a late southern summer morning humidity look, even though this was originally shot on a hot dry day.

The Dione AI models are some of the earliest models released in Video AI. There are no modifiable controls, aside from adjusting the field order or adding noise to counter some strong noise reduction and texture smoothing results.

General enhancement model for reducing noise and compression artifacts.

This model has been specifically trained on face recovery for low quality (LQ) and mid quality (MQ) input videos. Unlike the original trio of interlaced models, not only does Iris deinterlace footage, it gives the user creative freedom to modify video controls. This robust model can be used for more than just faces, as it does a good job of pulling fine details. 

Keep in mind that since this model is good at regenerating fine details, you will not need to be as aggressive with “improve details” in the manual settings. Additionally, this model can be used with both Interlaced and Progressive videos.

More details on this AI Model can be found in the Progressive section.

Proteus is a general enhancement model that allows fine-tuning several parameters for optimal results. The model excels with reducing noise when working with low to medium-quality footage. Proteus gives the user creative freedom to modify video controls in addition to deinterlacing footage. This model can be used with both Interlaced and Progressive videos.

More details on this AI model can be found in the Progressive section.

Interlaced Progressive are videos that were exported as Progressive, but have interlaced fields baked into the frames. Sometimes this could also mean that the field order was set incorrectly. Additionally, there is the possibility of compression or general softness if rendered out at a lower quality with baked-in interlaced fields.

These models should be used when working with interlaced footage that have been previously incorrectly transcoded to progressive.

This is a more aggressive deinterlacing model. This AI model was made specifically with Interlaced Progressive footage in mind.

Performs best on the following resolutions;

If results are not as expected, try one of the Interlaced Dione models.

Specializes in removing halo and fringing. Like Dione Robust, this is also a more aggressive deinterlacing model. This AI model was made specifically with Interlaced Progressive footage in mind.

Here is a comparison between Dione Robust and Dione Robust Dehalo.

Enhancement

Film grain is a unique attribute of analog media. The size, frequency, and characteristics of different film grain comes from the quality of the film stock itself.

When digitally applied, grain adds a visual texture that gives an organic feel.

Mimics the the look and feel of film, and can contribute to a more natural result when combined with AI model enhancement.

Adds a creative touch to otherwise clean digitally captured footage.

Grain is also used to mask banding patterns or low quality footage.

 is a fine digital sandy textured artifact that is a result of the image sensor.

 within the Enhancement section of the screen right Controls Sidebar. In older versions of the application, Grain is accessible just below “Motion Deblur” from the Controls Sidebar.

Click on the "Apply Grain" button to reveal the dropdown menu.

 controls how much visible grain will be applied to the footage.

 controls the scale of the grain texture.

Grain

Dynamic Range is the contrast between the brightest highlights and darkest shadows of a video. 

Generally 8-bits per channel with a total of 24 bits per pixel.

The brightest luminance value can range around 100 up to 400 nits (candela per square meter) depending on the display.

Has a total contrast ratio of around 1,200:1.

Holds a scope of 12-17.6 stops for video. Though most cameras can produce up to 15 stops.

Generally 10-bits per channel (but can have 32 bits floating point per channel) with a total max of 96 bits per pixel.

The brightest luminance value is around 10,000 nits (candela per square meter), but will only display up to 1,000 nits on most HDR monitors.

The maximum contrast ratio of around 20,000:1 on LCD displays and 1,080,000:1 on an OLED display.

Compared to SDR, HDR holds more range in the middle gray values, shadows and pushes the brightest and darkest values further apart. The limitation, really comes down to the viewing capability with our displays.

As noted above, in order to see HDR you will need to have a display that is capable of displaying these values. Luckily the most recent cell phones, TVs, desktop monitors and some handheld gaming displays can display HDR. Though, not all HDR displays are made the same. This is something you will need to do research on your own devices to know what’s the maximum values you can actually view.

If you view an HDR video on a non-HDR display, you may see clamped/flattened values. In the example above, the highlight values appear 'clipped' because they are out of range of the display. 

Image above has been photographed off of an HDR Display

Viewing the same clipped HDR image from an HDR monitor you are easily able to see a wider dynamic range, most notable in the clouds.

Above, you can see that the there is minor difference between the two images when viewed from an SDR display. Note the settings were slightly adjusted for the SDR to HDR tool.

Please take the two images above as a reference as they were photographed on an HDR display.

When viewing from an HDR display, the difference is a little more apparent. The values in the white water are more defined and not so flat, the contrast range has greatly expanded and the richness in color values has also improved.

The Topaz Video AI converts SDR to HDR by generating “Inverse Tone Mapping” (ITM). 

This process is also known as “Up Mapping”. This process uses the luminance values within the limited/clamped SDR imagery to “add” values or push more range to create a larger scope of color data. 

This also does a conversion of BT.1886/rec709 standard SDR content (which is how the color image is consistently seen between monitors and other digital displays) to BT.2100/PQ (Perceptual Quantizer) standard HDR, making the result viewable on an HDR screen.

The limited range originally created from a standard video can now be modified to have an overall wider range that requires preview on HDR displays.

HDR video has a closer range to what an average human eye can perceive compared to SDR video. 

You can find the SDR-to-HDR mode in the upper screen right side of the application within the “Control Sidebar”. This is between "Frame Interpolation" and “Stabilization”. Click in the sider button for the SDR to HDR to enable and reveal a drop down menu.

For the AI Model, clicking the carrot reveals a single AI Model. The “Hyperion - HDR (ALPHA)” is our model that does the inverse tone-mapping that increases color-depth, contrast and increased peak brightness.

Adjust Exposure allows the user to modify the exposure from the origin starting point, which defaults to “0”, up to double the max amount value of “1”. Note that the AI Model will intelligently choose the areas to brighten based on the video content.

Boost Saturation allows for an increase of color intensity from the original amount at “0” up to the maximum at “1”.

The Highlight Threshold slider sets the lowest intensity level in linear color that has been determined to be a highlight by the Hyperion AI Model. 

Lowering this value could cause color shifts & large contrast changes, especially in skin tones. It is recommended to test from the default setting of 0.70. If there are any slight color shifts, raise the value, otherwise you can experiment with lower ones. The Highlight Threshold is between 0.45 to 0.85.

In order to view any of the parameter changes, you will need to render a preview. With the updated player in Topaz AI 5.4, you will be able to view the HDR version of your video.

An HDR display will be needed in order to view the results properly.

At this moment the player is not updated for macOS, so the application will require an external player to preview the HDR updates.

QuickTime V210 Uncompressed YUV 10-bit 4:2:2

FFV1 does not play with QuickTime player. Not viewable in Premiere

VP9 does not pay with QuickTime player. Not viewable in Premiere

QuickTime Animation does not play with QuickTime player. Not viewable in Premiere

SDR to HDR is not available for the Adobe After Effects Plugin at this time.

SDR to HDR is not available for the Davinci Resolve Plugin at this time.

SDR to HDR

Functions (Legacy)

Quick Start (Legacy)

Understanding the filters and models within Topaz Video AI will allow you to create the best possible output. 

Filters (Legacy)