i usually like ffdshow's noise feature to enhance viewing, but in avisynth, i had no idea how to emulate it without reading the source code. so this may be way off, but it seems to be workable for adding grain/artificial detail, as the dshow mpeg-4 decode filters do. this would work ideally for low-medium bitrate "action" video (600-900kbps) to a higher bitrate end-product.

results are fast enough, and some noise is removed by the HybridFuPP() script, while sharpening the distinct edges and enhancing definition.

you may even want to limit the luma range in blockbuster to stop adding excess noise to bright areas, as ffdshow also allows for. depending on the source avi or final bitrate, this might be an advantage.

i didn't add Convolution3D(), TemporalSoften() or DCTFilter(1,1,1,1,1,1,0.5,0) from the original MA/DIKO script, as i wasnt in a need to post-process the unfilter/flux masks from within HyFupp, and im unsure as to the greater effect of DCTFilter.

any changes/suggestions are welcome, as i really only hacked this up in 4 hours of reading other posts and the scratchy online manuals. so very scratchy ...

ideally i'd move the order around, but it seems to work OK as it is, the pre-resize/HybridFupp() noise and the post-resize noise add definition to the edge sharpening mask, which will be softened by detailed noise, not a harsh blurring/ringing effect. i only really tried VagueDenoiser, which works too well at removing grain/detailed noise, but i didnt experiment much at all.

Why don't you let HybridFupp doing the resizing ?

And I would add that its NOT useful to add noise before HybridFupp as this wont end up in the same purposure like it does in the optimal scripts!
The purpose was to "spray" on existing artifacts/blocks so that they can be handled by for example ATC() and so in a much heavier temporal cleaning way.

And you do postprocessing using BlindPP before where BlindPP already is used in HybridFupp()

Re HybridFuPP resizing, i initially tried it, Avisynth threw up an exception error a few times. so i left it up to GripResize. if i knew how the mechanics of gripsize worked... i'd change it around to do bicubic -> lanczos, which is what preset="high" is set for.

after evaluating the HybridFupp v0.855alpha script deeper, i came to a conclusion, the filters are a little too strong on easily compressed footage (which is why i added noise). i tried to look at sky, smoke, foliage and skin, this seems to be more what i intended it to look like.

re: adding that initial BB noise might confuse the mask/edge detection (looking at Fupp's debug info, i couldn't notice any difference with with noise/edge mask displacement, just that more low-light areas were picked up by the 'green' edges with added noise), but it did so without changing parameters, which was easier at the time ...

anyway

And now you applied at the end a blockbuster followed by the noisekilling Vaguedenoiser and its spatial work 8O

What are you doing ??? :wink:

mainly two things, as hap-hazardly as possible. also, trying not to hijack the original thread, but the script doesnt do so well on low-quality bitrates unless you like your blurring, extra crispy.

i'm not used to reconstructing 600 kbps to 4mbps, and i'm doing this by feel, since documentation on avisynth filter systems is just like reading a german->japanese dictionary translated back to english by a snotty french waiter. ("sacre bleu, you should know this *sneer*")

mainly i want a script that will ...
convert a progressive, blurred NTSC 24ps to PAL 25fps, without using telecide or tomsmocomp to blend/process frames, and/or perform some frame/motion evaluation that AssumeFPS doesn't do (without having to do some wacky IVTC-field doubling analysis that will just bleed time and details with no effect. no offense to the people who deal with int fields daily, but it's far too late to fix it)

some motion/edge aware enhancement to increase detail and a detail sensitive noise/denoise process to add the detail back into the heavily DCT compressed blocks of the original file. (it's only because of some distant memory of wavelet image reconstruction in texture construction, that i even use it in preference to the normal spatio/temporal smoothers.)

i'm looking to emulate pretty much what ffdshow could do (overscan borders/resizer, luma-sensitive noise, adaptive post-processing with deringing/deblocking), but with a higher output quality. if all else fails, i could just wring an avs into ffvfw, then into CCE and do it that way.

i tried modifying the existing DIKO/MA/Optimal script, but i get too much detail reduction in the 8x8 blocks, especially in sky/smoke/water/skin areas, and it seems like avisynth can do a lot more to reconstruct the impression of image quality than to dismember, by squeezing YUV channel macroblock delta's, or by other measures.

Could you tell me more ? Are you using avisynth 2.55, masktools 1.4.15 and deen 1.0 beta 1 ?

Using a resizer before HybridFuPP is a nonsense...

If I remember well, 0.855 uses deen("c3d",1,2,3,4,6,0.3,9) as a default for N2 parameter, which is a little bit too strong.

You would rather like to use hybridFuPP(preset = "high", N2="""deen("c3d",1,2,3,2,3,0.3,9)""") as a workaround or directly hybridFuPP(preset = "high", N2="VagueDenoiser(threshold=0.3,method=1,nsteps=6, chroma=true)"). Be aware that using N2, filtering is applied after resizing. If you want to apply it before, you should use N1.

PS : Blockbuster shouldn't be necessary after HybridFuPP

Regards,
FuPP

yes to all. i hunted down an old/distinct beta copy of deen, and that works fine, iirc, the avisynth exception came up if i used HybridFuPP(720,576,preset="high"), but not if i used width=720, etc. it might also be related to avisynth memory alloc, but with less filters loaded, it works fine.

ive only recently become more interested in looking at masked filters to only apply an effect to needed areas, since most destroy fine detail unless used sparingly.

i'd ditch GripResize, if i could, but it's so much easier ...
for N2, deen("c3d",1,7,9,4,6,0.5,0) is what im looking at in the original avs. i'll make the change and see how it goes.

it worked great. after a lot of unnecessary wailing, its fine.

i modified my script a little after trying to reduce the HybridFupp preset de-noise thresholds, and moved from noise to 'dither' in Blockbuster. the quality looks and seems to be just like the original source, with no discernable blocking, constrasts and edges are distinct, looks fine.

except for it being 12-15fps. at best. a different denosier would work a lot faster. also using scripting instead of GripResize to do anamorphic ratio sizing /overscan, would be a plus.

That's no good at all :!:
Even SansGrip agrees, and the dither option was experimental.

-kwag

Did you try without blockbuster (actually, I'm not sure to understand why you use it) ?

I definitively do not like the idea to resize twice... It must slow down terribly things !

You can use a simple crop() and addborders() and help you with fitcd or something else to do the calculations...

Btw, I will have a look to see if I can do something with SansGrip thingies and HybridFuPP...

FuPP

ok, i feel silly. works better without Blockbuster, its now unnecessary.

all this ended up because of harsh denoising in presets which i never really looked at to begin with. as the saying goes .. less, is more :)

all of the help is very much appreciated.

and Fupp, if you do get code to work overscan borders in, as a resize option in some future HybridFupp(), i'd expect a lot of old and new fans in KVCD would appreciate it too.

ok, I've just discovered that doing the following, you do not need Gripsize anymore :


Regards,
FuPP