Resolution Upscaling & DeNoise

Resolution upscaling allows you to upscale videos from, say, 1080p to 4K. Upscaling videos is provided using either Waifu2x or SRMD. Both use NCNN as the graphics backend which is currently the fastest implementation for them. In Waifu2x, you can also configure the settings to DeNoise only without upscaling.

Usage

Selecting a Tool

Waifu2x (NCNN):
Provides up to 2x upscaling, hence the name. It’s one of the fastest AI resolution upscaling tools there is currently. DeNoising is also built into Waifu2x which you can use simultaneously while upscaling or as a stand-alone feature. Although Waifu2x is fast, it is lower quality than the SRMD alternative.

SRMD (NCNN):
This is an alternative upscaling option to Waifu2x which allows up to 4x upscaling, but the rendering speed can be about 2x slower than Waifu2x. DeNoising can also be used simultaneously with SRMD. SRMD has better quality upscaling than Waifu2x.

Selecting a GPU Device

Allows you to select a GPU device to use for upscaling. Select a Vulkan capable GPU in the list.

Selecting GPU Threads

Here, you can set the number of GPU threads for the upscaling process to use. An all around good value to use is 2 threads. More threads will not make the process faster and can throw errors if too many are used.

Selecting AI Model (Waifu2x Only)

This section allows you to select an AI model for upscaling.

Upconv 7 Anime:
This model is designed for anime films or 3D cartoons.

Upconv 7 Photo:
This model is designed for reality/film videos. Don't use this model for 2D videos.

Cunet:
This model is specifically designed for 2D videos. However, this model is also known to run 2x slower than the other models above.

DeNoise Only: This "model" allows you to only DeNoise the video without upscaling. It uses the Cunet model but without upscaling.

Selecting Noise Reduction Level

This option allows you to set the noise reduction level for the video. If your video has no noticeable noise, then you'll want to set this value to Medium in Waifu2x or 3-4 in SRMD. If your video has noticable noise, then you'll want to set this value to a higher selection. Finding the right noise reduction value for an individual video can take some time, so please be patient as you test different values.

Selecting Color Precision (Waifu2x Only)

Color precision specifies the accuracy of the colors during the upscaling process.

Fp32:
This is the recommended option. Fp32 is a high color precision level which allows a higher quality upscaling at the cost of a slight rendering speed loss.

Fp16:
This option is faster than Fp32 and is suggested to use if you want a faster upscaling process and don't care so much about the quality of the colors in the video.

Using TTA (SRMD Only)

Test Time Augmentation (TTA) allows the upscaling to be more quality focused at the huge cost of very slow rendering speeds. Since using TTA is so slow and only has slight quality increases, it is generally not recommended to use.

Selecting Output Resolution

Lastly is to set the output resolution. Keep in mind of the resolution upscaling limits in both Waifu2x and SRMD. Waifu2x only supports a maximum resolution of 2x the source video resolution whereas SRMD supports up to 4x. The default output resolution is set to maximum output resolution possible.

Auto Adjust:
This setting allows EncodeGUI to automatically adjust either the output resolution width or height on the basis of user defined width or height in accordance with the video aspect ratio. You can disable this option so that you can set the width and height to be outside of the source video's aspect ratio. However, keep in mind that if the custom upscaling resolution you select consists of a different aspect ratio from the source, then the video will be stretched.

Dual GPU Acceleration

This feature can be configured in Preferences -> Encoding. This function allows you to utilize two dedicated GPU devices to increase upscaling speeds for Waifu2x or SRMD by up to two times. The first selected GPU will do the even frames, whereas the second GPU will do the odd frames. It is not ideal to use this tool with two largely different speed GPUs as the both GPUs will process the frames at the speed of the slowest GPU of the two selected. This is because both the even and odd frames must be processed in even synchronization.