On Thu, Dec 13, 2012 at 7:08 PM, Brian Matherly <[email protected]> wrote: >>>> Yes, there are significant quality issues. The composite bounding > >>>> geometry being used here is set for 100% = the size of the profile. >>>> So, I doubt that it is scaling. However, if you use a non-square pixel >>>> profile, then the title's is going to be scaled horizontally to >>>> conform its square pixels to the profile's sample aspect ratio. If >> you >>>> combine that with the sdl consumer's default scaler setting of >> nearest >>>> neighbor, you will get poor quality. However, in my brief tests I >>>> specifically avoided these pitfalls. I even found the problem when >>>> using the Qt-based kdenlive titler. Certainly, the combination of the >>>> colors chosen and SD resolution limit the quality, but I noticed that >>>> the affine transition's compositor does give better results. The >>>> regular composite transition uses yuv422 with its reduced chroma >>>> resolution, and that at least partially contributes to the problem. >>>> Perhaps when you combine that with composite's lack of >> interpolation >>>> and these colors and low resolution. >>> >>> I produced these two images during my analysis. They were created in >>> kdenlive using the kdenlive titler (not dynamictext's pango). They >>> were made using a 768x576 square pixel profile. The right side of the >>> image within the gray box does NOT use any of mlt's compositors; it >>> uses Qt's. It is simply text over a rectangle were the rectangle is >>> filled with the same color as the background color clip. The left side >>> demonstrates the MLT transition used - affine vs. composite. If you >>> flip back and forth between the two within an image viewer, it shows: >>> >>> a) The affine composite quality is roughly the same as Qt's (look at >>> 'S' and 's'). >>> b) The composite transition appears to have an off-by-one error >>> between the image and the alpha channel. I do not yet know if it is >>> really off or if it is just side-effect of the simple implementation. >>> c) There is a lot of aliasing even within the Qt composited area as a >>> result of low resolution. >> >> The previous two images were generated by rendering to H.264 and then >> converting the first frame to PNG. Here are two new ones that render >> straight from the project to PNG. > > Based on those results, is there any reason the watermark filter shouldn't > use the affine transition by defaut?
There are subtle differences between composite and affine's simple compositing use cases and not so subtle differences in performance (composite is faster). So, I do not want to suddenly change behavior of watermark for people who are already happily using it. I would like to see if composite can be improved/fixed, however. But I do think you should consider to change dynamictext to not use watermark and instead use affine through either filter_transition or code copied from watermark. Also, you could write a Qt backend to dynamictext that writes directly on frame instead of using an external composite. qimage/transition_vqm shows how to do that. -- +-DRD-+ ------------------------------------------------------------------------------ LogMeIn Rescue: Anywhere, Anytime Remote support for IT. Free Trial Remotely access PCs and mobile devices and provide instant support Improve your efficiency, and focus on delivering more value-add services Discover what IT Professionals Know. Rescue delivers http://p.sf.net/sfu/logmein_12329d2d _______________________________________________ Mlt-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/mlt-devel
