Spaces:
Running
on
A10G
A little suggestion
@dvruette
, hello, it is me again. Nothing is broken (well, your other space is, you maybe want to look into it), but I was using this one (fabric) for a while, and only recently came to realization (maybe an obvious one), that negative scale actually works per individual picture, rather than entire group as I thought. That is, if I have two pictures in liked group and three in disliked, I have to set scale to 2/3 (rounds to 0,67 but it's a small error margin), otherwise quality of resulting image will tilt. But then, if I have three liked images and two disliked, I cannot set scale to 3/2 (1,5), since 1,0 is maximum. Can you plese tweak it around, so that I can use it this way, rather that add copies of pictures to even them out (which is technically unacceptable, as it increases drawing time, and system load; in the example above, adding copy of negative pictures adds two additional pictures, so that scale is 3/4 (0,75), which is 7 pictures total, instead of 5, and 7 pictures is pretty heavy load.)
I understand that locking option to always be this way may upset those, who want image slightly out of focus (and those who have artistic eye; I don't), but can you increase upper limit to, say 5, or 10, maybe? Many thanks in advance.
(I understand that you maybe have it in a released version, that I saw recently, and this is but a demo, but my computer is kinda old, and so I'm stuck with this. This one is still very good thing, although I keep coming up with new techniques for continuous progression of quality, because older ones keep ending up with a pretty weird results...)
That's quite interesting, so I assume you're talking about the "Neg. feedback scale" slider in the feedback parameters section? It never occurred to me that one might want to put more weight on negative images than positive ones, but there's no particular reason you shouldn't!
It also should be a simple fix, I'll look into it.
Maybe I'm thinking too mathematically, but I think like this: let's say I want to combine two pictures. If I just put them together, their internal values will combine and overflow. (If I reduce the feedback scale itself, then I have to guess whether the AI will combine the right features or not). So I place a counter image to negate the effect. And then I noticed that one image doesn't do the trick, but two images do. But then it's possible that the AI, using two images, will subtract some features that I want. So two copies of the same image will work. But why should I use two copies when I can just put more weight on the negative image itself? By the way, maybe it's also reasonable to increase the normal feedback scale to increase the initial values.