I've always wondered something about this. Why compress the media itself before archiving instead of just using archive compression like 7zip's "compress and archive" option?
7zip and lossless compression algorithms can't beat in ANY way lossy compressions, but this is at the cost of overall quality, of course. In particular, generic lossless algorithms near can't compress at all media (video and audio in particular, uncompressed raw images can sometimes be a bit compressed).
So, you compress all media (audio, video, pictures) with high compression lossy algorithms, deflating the size by an HEAVY factor (virtually unlimited, you're stuck only with final quality in fact), then you use a lossless compression algorithm (LZMA2 for example) for program code (Ren'Py's program and binary libraries, Python scripts, archives, etc.) to gain a bit more space.
But lossy compressions can lead to strange things: first, on image-based animations, compression artifacts (quality loss) can be different from one image to another, resulting on screen by strange color glitches and/or deformations during animation. In all cases, you often lose texture details. If images are JPEG in original game, a double compression often produces a bad result: they should be PNG format for best result.
For videos, dimension can be smaller (then artificially resized on replay), which statistically cut the size by 4 when you cut resolution by 2, and/or use a more aggressive compression algorithm, which can cut size by near any factor. But since most videos in such games are already compressed, a double compression often GREATLY affects quality.
Finally, audio: same as video, but strangely, human ear is more sensible to audio defects than to visual defects. A common compression artifact is to have a "metallic" sound in the final result, with what seems to be a bit higher pitch in voices.
On overall: compressed games are a good thing when you can't afford to waste bandwidth for just testing a game. But if you're not in a country where Internet usage is paid on a data volume basis, skip these versions and stick to the original instead.