[FFmpeg-devel] Lossy GIF encoding

[Tomas Härdin]

fre 2019-02-15 klockan 09:01 +0000 skrev Kornel:
> Hello!
> 
> I'm working on implementing lossy gif encoding, and I'd like to hear
> your opinion on how to best integrate it with ffmpeg in a way that
> would be acceptable for inclusion in the project.
> 
> Lossy encoding is done by making LZW use approximate matching, and
> with the right input it can halve GIF file sizes. I've previously
> implemented it for gifsicle: https://kornel.ski/lossygif
> 
> The main problem I have now is that the existing GIF codec receives
> frames already converted to AV_PIX_FMT_RGB8 or similar 8-bit format.
> When working with 8bpp heavily-dithered input, the lossy compression
> struggles to find runs of similar pixels, so it isn't very effective.
> 
> For proper, high-quality lossy encoding I would need access full-
> quality input, ideally something like 24/32-bit RGB/A. However, I
> don't know if it's possible to do that within the existing GIF codec
> implementation.
> 
> libavcodec/gif.c in ff_gif_encoder.pix_fmts seems to passively
> declare types of pixel formats it accepts. Is this list fixed? Could
> it be made varying depending on flags? I'm thinking that if user sets
> a flag that opts in into lossy encoding, the codec should request
> AV_PIX_FMT_RGBA input instead of the 8-bit inputs (or both at the
> same time, if that's possible).
> 
> Alternatively, would it be OK to create a separate codec just for
> lossy compression? I would create a new GIF codec, define it under a
> new name, and declare it always takes AV_PIX_FMT_RGB24 or
> AV_PIX_FMT_RGBA. Is that a good approach?

This sounds more like something that belongs in libswscale (RGBA ->
PAL8 conversion) or could be done with a filter. An option like
ImageMagick's -fuzz could be added to the PAL8 quantizer in sws

I could mention that I have a patch for the palettegen filter that I
haven't pushet yet

/Tomas