OK so I'll start to address some of the things needed to be done to get
a good, solid 4.8 release. I'll spread this across several messages as I
get time.

(Before I begin: I have yet to fully catch up with this thread as there
are a number of messages I've left unread - I'll catch up later,
apologies if I duplicate/repeat something already covered).

One of the main issues we face (and where we could really do with
community help as this is hard to automate) is making sure all the
combined changes firstly don't outright break any of the shipped scenes
and snippets (by that I mean cause parse/render to fail with an error or
exception), and secondly don't cause any unexpected *visual* artifacts
(or at least not any unexpected ones that would fail the 'principle of
least astonishment' test that I ramble on about occasionally).

The first can be automated to a reasonable extent. The latter is really
part of the former (as to view the image you need to complete the
render) and is really, really difficult to automate. The reason for that
is it's a matter of perception; of a human's ability to look at
something and see it for what it is rather than a collection of pixels.

Computers are OK at looking a collections of pixels mathematically if
they have a base set of images (rendered e.g. by v3.7) to compare
against. They can see if a pixel is different by 0.001% and can even be
taught to understand things like gamma changes but it's much more
difficult to get an algorithm that correctly interprets what a *user*
would consider a 'right' or 'wrong' change from an image as a whole.

It really needs user input. If we could get a cadre of all the above-
mentioned images rendered with 3.7 (plus a selection of good external
candidates like HOF images and some newer works), then the same with
3.8, and have them all put up side-by-side on a website or some other
means of allowing multiple collaborators to view, rate and/or comment on
the differences then it would be a large step forward to getting 3.8 out
of the door.

-- Chris

Post a reply to this message