do you recommend any good source to read? i was thinking about creating sum table for the whole image
getting variable size area average for each pixel would have just 4 lookups (and all parallel) regarding quality i was thinking doing it (box blur) 3x to approximate gaussian sum table should be possible to generate in parallel (for one dimension at a time) M On 15 Aug 2010, at 19:10, vade wrote: > Read up on separability and convolution, and multipass tricks. This can get > much much much faster and maintain quality. > > On Aug 15, 2010, at 8:31 AM, Miroslaw Rusin wrote: > >> thanks for the info. so it's something more like this then... >> >> <compound-3.qtz> >> >> >> it's slow - O(x*y*r²), where: >> x - image with >> y - image height >> r - radius >> >> can be O(x*y) with sum table >> >> M >> >> On 13 Aug 2010, at 22:29, Patrick Sheffield wrote: >> >>> I was about to write the same email (thank you Vade). Thank you Miroslaw, >>> but yes - what you're doing is more of a diffusion - blending between a >>> blurred and clean image... >>> >>> I think I get it now. The degree of blur is determined by how much a pixel >>> is averaged with its surrounding pixels. If you control that with a value >>> from a mask image, you will have a compound blur... >>> >>> Patrick >>> >>> >>> On Aug 13, 2010, at 1:12 PM, vade wrote: >>> >>>> Blending between a blurred and un-blurred image does not produce a >>>> variable width blur. You would need n number of blurred images to do that, >>>> where n number of images. >>>> >>>> What you do is blur a variable amount. : >>>> >>>> Note the difference in the top blur (our compound blur) >>>> >>>> and the mixing between a blurred and non blurred image >>>> >>>> http://i.imgur.com/TjCZY.png >>>> >>>> http://vade.info/CompoundBlur.mov >>>> >>>> Im sorry but I cannot disclose our algorithm, but if you research how to >>>> do a fixed width convolution kernel blur (box or gaussian), separate it >>>> (for speed and cleanliness) it should be apparent how an additional >>>> sampler input can modulate a uniform float input :) >>>> >>>> As for the note about mip mapping, that is certainly a fast approach, and >>>> mip-mapping is possible within QC if you use the Image Texturing >>>> Properties patch. Note howeveer that forces GL_TEXTURE_2D (as rect is >>>> default in core image/video and Quartz Composer), can mean some patches >>>> act very odd with an image input thats 2D (or flat out fail if its a 3rd >>>> party QC input and they dont check the [self.inputImage textureTarget] >>>> value when binding. Just saying, it act odd, but really should not :) >>>> >>>> On Aug 13, 2010, at 2:08 PM, Miroslaw Rusin wrote: >>>> >>>> >>>> >>>>> what's the algorithm for compound blur then? >>>>> >>>>> attached demo with example pics from >>>>> http://livedocs.adobe.com/en_US/AfterEffects/8.0/help.html?content=WS3878526689cb91655866c1103a9d3c597-7baa.html >>>>> >>>>> top-right corner picture is generated >>>>> >>>>> all in "Compound Blur" macro if you want to copy-paste in your composition >>>>> >>>>> also added blurred img cropping; you can change gaussian blur to >>>>> something else if you like as well >>>>> >>>>> M >>>>> >>>>> <compound-2.qtz> >>>>> >>>>> On 13 Aug 2010, at 16:28, Stonewall Ballard wrote: >>>>> >>>>>> A constant blur with a modulated blend produces a different effect than >>>>>> a true compound blur. I don't know about doing this in QC, but for other >>>>>> work, I use a mip-map or a summed area table. Can you get at a mip-map >>>>>> in QC? >>>>>> >>>>>> - Stoney >>>>>> >>>>>> On Aug 12, 2010, at 2:48 PM, vade wrote: >>>>>> >>>>>>> Thats totally not rendering correct for me, nor is it the best way of >>>>>>> handling it. You can do it in a single custom core image filter. Im not >>>>>>> able to hand out code, but read up on separable convolution kernels, >>>>>>> and get it working for a constant amount for the entire image, and then >>>>>>> modulate that amount by the intensity (a dot product) of a second >>>>>>> sampler. Also you will need to pay attention to destination coordinates >>>>>>> in core image, ie, ROI and DOD since you are handling multiple input >>>>>>> images and samplerCoord may not be correct depending on your GPU, and >>>>>>> size of input image. >>>>>>> >>>>>>> On Aug 12, 2010, at 2:11 PM, Miroslaw Rusin wrote: >>>>>>> >>>>>>>> >>>>>>>> <compound.qtz> >>>>>>>> >>>>>>>> <Screen shot 2010-08-12 at 20.09.49.png> >>>>>>>> >>>>>>>>> >>>>>>>>> There are often times where I want to blur an image based on a mask >>>>>>>>> and simply using the mask to blend a blurred and a clean version of >>>>>>>>> the image is not enough - I want to blur the image more or less, >>>>>>>>> based on the intensity of the mask. Performance always matters, but >>>>>>>>> quality is more important... HD in realtime is probably not doable, >>>>>>>>> I'm thinking... so processing and rendering will be acceptable levels >>>>>>>>> of performance... >>>>>>>>> >>>>>>>>> Patrick >>>>>>>>> >>>>>>>>> P.S. - I did mean Quartz Composer >>>>>> >>>>>> -- >>>>>> Stonewall Ballard >>>>>> [email protected] http://stoney.sb.org/ >>>>>> >>>>> >>>>> _______________________________________________ >>>>> Do not post admin requests to the list. They will be ignored. >>>>> Quartzcomposer-dev mailing list ([email protected]) >>>>> Help/Unsubscribe/Update your Subscription: >>>>> http://lists.apple.com/mailman/options/quartzcomposer-dev/doktorp%40mac.com >>>>> >>>>> This email sent to [email protected] >>>> >>>> _______________________________________________ >>>> Do not post admin requests to the list. They will be ignored. >>>> Quartzcomposer-dev mailing list ([email protected]) >>>> Help/Unsubscribe/Update your Subscription: >>>> http://lists.apple.com/mailman/options/quartzcomposer-dev/psheffield%40earthlink.net >>>> >>>> This email sent to [email protected] >>> >> >> _______________________________________________ >> Do not post admin requests to the list. They will be ignored. >> Quartzcomposer-dev mailing list ([email protected]) >> Help/Unsubscribe/Update your Subscription: >> http://lists.apple.com/mailman/options/quartzcomposer-dev/doktorp%40mac.com >> >> This email sent to [email protected] >
_______________________________________________ Do not post admin requests to the list. They will be ignored. Quartzcomposer-dev mailing list ([email protected]) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/quartzcomposer-dev/archive%40mail-archive.com This email sent to [email protected]
