Keyframes are meant to be visible for less than 100 milliseconds in typical cases. Not to say that AV1 is bad at still images, but there are tradeoffs associated to this fact ("low fidelity imagery").
> Keyframes don't disappear with the next frame — they're the basis of all the delta (P- and B-) frames that follow until the next I-frame.
Exactly this. P/B-frames are not encoded as a delta to the input I-frame; it is encoded as a delta to once compressed then decoded I-frame, otherwise errors in the I-frame will accumulate. So any imperfection in I-frames has a chance to be fixed by subsequent P/B-frames, but this is impossible in a still image.
You have a good point that AVIF layered image items can act like such P/B-frames. Do libavif (or other AVIF implementations if any) make use of them? JPEG XL has the same feature (zero-duration frame) but I think libjxl doesn't use it to encode a residue when the input image is not animated---in fairness such residues are very hard to compress though.
> You have a good point that AVIF layered image items can act like such P/B-frames. Do libavif (or other AVIF implementations if any) make use of them?
