DeGrainMedian - FFT3DFilter - GoldDust - HQDN3D - LRemoveDust - PixieDust

( Home - Spatial Smoothers - Temporal Smoothers - Sharpeners - Dealing with Dot Crawl - Dealing with Rainbows )

• limitY (0-255) determines the maximum amount by which DeGrainMedian may change a pixel's luma value. If the luma would have been changed by more than this value, it is changed by exactly this value instead. Higher values will increase the strength of the denoising (but going too high may cause overblurring).
• limitUV (0-255) determines the maximum amount by which DeGrainMedian may change a pixel's chroma value. Everything just said about limitY also applies here.
• mode (0-5) determines the filter's operation mode; all you need to know about it is that higher values will weaken the smoothing. The higher you set mode, the higher you can usually safely set limitY and limitUV.
• interlaced (true/false) should be set to true if your source is interlaced.

• DeGrainMedian should be applied before other denoisers.
• DeGrainMedian can be stacked twice to good effect if the limits are low enough.
• limitUV can usually be safely set to higher values than limitY.
• If DeGrainMedian is oversmoothing your image, try a higher value of mode; if that doesn't work, reduce your limitY and/or limitUV.
• Obviously, if your source is interlaced, be sure to set interlaced=true.

Defaults: FFT3DFilter(sigma=2.0, plane=0, bw=48, bh=48, bt=3, ow=16, oh=16, sharpen=0, interlaced=false)
Parameters:

• sigma (any float > 0 ) indicates the noise level of the video. Increasing this value will strengthen the denoising but may introduce oversmoothing or gridding artifacts.
• plane (0-4) determines what color planes the filter should process:
  • plane=0 processes only the luma.
  • plane=1 processes only the U chroma channel (the blue offset).
  • plane=2 processes only the V chroma channel (the red offset).
  • plane=3 processes both chroma channels.
  • plane=4 processes all three channels (luma and chroma).
• bw and bh (any positive integers) determine the width and height respectively of the blocks to be used for the transforms. Lowering these values will speed up the filter but may make gridding artifacts more likely; raising them may improve denoising but will slow down the filter. Setting these too high or too low may make the filter useless; Fizick recommends values from about 32 to 64.
• bt (-1 to 4) determines the nature of the temporal element of the filtering:
  • bt=-1 does sharpening only without denoising, 2D only. (If you use this mode, make sure to specify a value for the sharpen parameter; otherwise the filter will do nothing.)
  • bt=0 does Kalman filtering that takes into account all previous frames, allowing for stronger noise reduction than the Wiener modes.
  • bt=1 does 2D Wiener filtering (spatial only).
  • bt=2 does 3D Wiener filtering taking into account the previous frame.
  • bt=3 does 3D Wiener filtering taking into account the previous frame and the next frame.
  • bt=4 does 3D Wiener filtering taking into account the two previous frames and the next frame.
• ow and oh (any positive integers less than half of bw and bh respectively) determine the overlap width and height for each block. Larger values will reduce gridding/blockiness/ringing but will slow down the filter.
• sharpen (any float?) determines sharpening strength; by default this is set to 0, for no sharpening at all. Fizick recommends values of about 0.3 to 1.0 if you want to sharpen; anything higher than about 2 is probably a bad idea. Negative values will actually have the opposite effect, softening your edges (somewhat similar to UnFilter's behaviour).
• interlaced (true/false) should be set to true if your source video is interlaced.

Usage notes:

• If you encounter gridding artifacts, then try one or more of the following:
  • Increase ow and/or oh
  • Increase bw and/or bh
  • Decrease sigma
• FFT3DFilter can also do sharpening in conjunction with denoising; if you wish to do so, set sharpen to something higher than 0 (but keep in mind that this will counteract your compression gains somewhat). If you only want it to do sharpening (no denoising), then set bt=-1 as well.
• Of course, if you encounter oversmoothing, then reduce sigma.

I originally didn't think I would be posting any comparison frames on this page, but I found one for which the effect was so striking that I simply had to. It's from Fushigi Yuugi: Eikou Den, episode 1, as filtered with FFT3DFilter(sigma=3,bw=32,bh=32,ow=16,oh=16,plane=4) (hover your mouse over the image to see the filtered version):

I think you'll agree that the difference is quite stunning; pick almost any mosquito noise on the original frame -- there's plenty to choose from -- and watch it disappear in the filtered frame. Higher sigma values will pound the noise even further into the ground, but often at the cost of harming other scenes by softening details. Back to top

• limit (1-??) determines the maximum amount GoldDust can change each pixel by; if a pixel would have been changed by more than this value, it is changed by exactly this value instead. Raising this value will increase the strength of the smoothing (but may hurt details).

• IMPORTANT: The Dust filters do NOT support YV12. This is because they were written for AVISynth 2.0.x, which was incapable of working in the YV12 colorspace; a version for AVISynth 2.5.x was never coded. If you try to feed GoldDust YV12 input, your application will suddenly quit without warning. Always convert to YUY2 first.
• IMPORTANT: The Dust filters do not autoload. Again, this is because they were written for AVISynth 2.0.x and not 2.5.x. To use them, you first have to load a plugin called LoadPluginEx2.dll (in a separate directory from your default plugins directory, because it will break autoloading otherwise); I've taken care of this for you in an autoloading .AVSI helper script. Then you manually load DustV5.dll by putting the following line before calling the filter: LoadPlugin("C:\avs25\plugins\DustV5.dll") (of course, replace the path given with the path to your AVISynth plugins directory).
• IMPORTANT: Do not call the Dust filters more than once in one script. This also means that if you use GoldDust in a script, you can't use SpaceDust, FaeryDust, or PixieDust in that same script.
• GoldDust should probably be used after cropping and resizing, but I'm not sure about this.

input=last.ConvertToYV12()
LoadPlugin("C:\avs25\plugins\DustV5.dll")
ConvertToYUY2().GoldDust().ConvertToYV12()
Repair(last,input)

The number one complaint about the Dust filters is blockiness; Repair may help some with this, but if you get this problem and you can't alleviate it, then you should probably switch to a different filter. Back to top

Defaults: HQDN3D(ls=4.0, cs=3.0, lt=6.0, ct=4.5)
Parameters:

• ls and cs (0-255 float) are the strengths of the spatial filtering on the luma and chroma respectively. Increasing these values will improve the smoothing but may overblur; anything above about 10 is probably not a good idea.
• lt and ct (0-255 float) are the strengths of the temporal filtering on the luma and chroma respectively. Increasing these values will improve the smoothing but may cause ghosting; anything above about 13 is probably not a good idea.

Usage notes:

• If your image looks oversmoothed (watch especially for chroma bleeding!), reduce ls and cs; if you notice ghosting/frame blending, reduce lt and ct. That's all there is to it.

Defaults: LRemoveDust_YV12(clmode=17, limit=2)
Parameters:

• clmode (0-18) determines what mode to use for the call to RemoveGrain in the script; I list all 18 modes here, and I'm not going to do it again. Don't use modes 13-16 for this.
• limit (0-???) determines the limit parameter to use for the call to LimitChange in the script; I have no idea what this does, because I don't know where to get documentation for LimitChange (it's not included in the RemoveGrain documentation, unfortunately). It probably limits how much each pixel can be changed by the function, but don't quote me. Thankfully, the default value of 2 seems to work well enough.

Usage notes:

LRemoveDust, as far as I can tell, stands for "Limited RemoveDust", Didée's modification of Kassandro's original RemoveDust script (found in the RemoveGrain documentation); it works by:

• Applying Clense [sic], basically a version of RemoveDirt (not to be confused with RemoveGrain) that offers NO protection against artifacts.
• Fixing the artifacts from the first step with Repair.
• Applying RemoveGrain to the repaired video.
• Applying LimitChange to the result (compared to the original input video); I believe this simply limits how much each pixel can change from the original to the final result, hence the "limited" part of the name.

• limit (1-8?) determines the maximum amount PixieDust can change each pixel by; if a pixel would have been changed by more than this value, it is changed by exactly this value instead. Raising this value will increase the strength of the smoothing (but may hurt details).

• IMPORTANT: The Dust filters do NOT support YV12. This is because they were written for AVISynth 2.0.x, which was incapable of working in the YV12 colorspace; a version for AVISynth 2.5.x was never coded. If you try to feed PixieDust YV12 input, your application will suddenly quit without warning. Always convert to YUY2 first.
• IMPORTANT: The Dust filters do not autoload. Again, this is because they were written for AVISynth 2.0.x and not 2.5.x. To use them, you first have to load a plugin called LoadPluginEx2.dll (in a separate directory from your default plugins directory, because it will break autoloading otherwise); I've taken care of this for you in an autoloading .AVSI helper script. Then you manually load DustV5.dll by putting the following line before calling the filter: LoadPlugin("C:\avs25\plugins\DustV5.dll") (of course, replace the path given with the path to your AVISynth plugins directory).
• IMPORTANT: Do not call the Dust filters more than once in one script. This also means that if you use PixieDust in a script, you can't use SpaceDust, FaeryDust, or GoldDust in that same script.
• PixieDust should probably be used after cropping and resizing, but I'm not sure about this.

As with the other Dusts (except SpaceDust), all I know about its theory of operation is that it uses some kind of motion compensation. According to Steady, it should provide the maximum compression gains of any of the Dust filters; it certainly does usually provide awesome compressibility gains. Its two major downsides (for both of which it is notorious) are its painfully slow speed and the blocky noise it often causes. However, such blockiness can often be mitigated through use of the Repair filter, and in such a manner as this:

input=last.ConvertToYV12()
LoadPlugin("C:\avs25\plugins\DustV5.dll")
ConvertToYUY2().PixieDust().ConvertToYV12()
Repair(last,input)

Something to watch out for: Here you can find a screenshot from my FLCL video "My Evil Alien" after having been run through PixieDust. This frame is taken from the middle of a fade in from black; it's on such fades to/from black that I usually find PixieDust to cause the most blocking. I mean, look at the sky. Uuuuugly. But on the plus side, the file size is nice and low. Back to top