Hi @DarthSim, sorry for breaking this :( I'll look at your PR tomorrow.

Hi again,

I noticed a failure with skate.gif. With:

$ vips copy skate.gif[n=10] x.gif

I see:

I think it's just because with this GIF there are no transparent palette entries. If I change the transparancy section to this:

        /* Pixels which are equal to pixels in the previous frame can be made
         * transparent, provided no alpha channel constraint is present.
         */
        if( page_index > 0 &&
                has_transparency &&
                !has_alpha_constraint ) {
                vips_foreign_save_cgif_set_transparent( cgif,
                        cgif->previous_frame, frame_bytes, cgif->index, 
                        n_pels, 0 );

                frame_config.attrFlags &= ~CGIF_FRAME_ATTR_HAS_ALPHA;
                frame_config.attrFlags |= CGIF_FRAME_ATTR_HAS_SET_TRANS;
                frame_config.transIndex = 0;
        }
        else {
                /* Take a copy of the RGBA frame.
                 */
                memcpy( cgif->previous_frame, frame_bytes, 4 * n_pels );
        }

(ie. only set pixels transparent if we have transparency in the palette) it seem to work.

Of course this means that GIFs which are video clips will hardly ever use transparency optimisation, since they will usually fill the palette and not have a transparent entry.

Perhaps vips__quantise_image_quantize_fixed() should always make palettes with index 0 as transparent?

We have a little code in vips_foreign_save_cgif_build() that tries to do this for the first frame, but of course that's often not enough.

Ah, @kleisauke suggests another fix: #2852 (comment)

        vips__quantise_set_max_colors( cgif->attr, (1 << cgif->bitdepth) - 1 );

To leave a palette entry clear for transparency. We'd still need to add the transparent element though.

Ah of course that's also in @dloebl's PR: #2863 ahem I'm slowly catching up.

... so combined with that other PR, all my test GIFs pass. Thanks you for fixing my mess, @DarthSim.