William Starr: How does one create collages like: https://www.facebook.com/photo.php?fbid=1637697329610635&set=oa.147858726033302&type=3&theater&ifg=1 So many people have them that it seems to me that there’s an easy mechanism for it, but if there is I don’t know the secret.
Bob Turner: I would start with a program like photoshop. I think the point of making one is to get the colors and patterns in the areas where the source photo would benefit. I’m thinking of the upside down girl with blue on the top, would benefit a landscape picture with a blue sky.
Daniel Prust: It doesn’t seem to matter where the colors physically are in the source image when compared to the style image collage, the AI will translate them to the right spot. But the point IS to create a style image with enough range in color, light, and texture so that the AI has enough to work with as to proportionally translate them to the appropriate sections of the source image. It will however translate texture directionally. If that makes sense😂
Ben Beekman: Daniel Prust I agree with most of this— however it’s not just texture but also general shapes that are affected by the orientation of the style image. Leading lines and color relative positioning were both affected in my test with rotating a rainbow style image in addition to texture orientation.
Style image rotation and its effect on dream output
Daniel Prust: Awesome, thanks for doing that! What I was saying is that it doesn’t matter that red is only found in the top half of the style image, or that blue is only found in the bottom half. It will translate the colors to all ‘reddish’ or ‘blueish’ parts of the source image in the formation, direction, shape, or “texture” of the style image.
Ben Beekman: That is certainly true, what I hope that illustrated is that it will paint predominantly with lines in the same orientation as they appear in the style— high levels of enhance probably give them more flexibility in this regard.
Andrew Irving: Hi Daniel, yes works by DD identifying a number of variables in its analysis of both the content and style image (they are analysed separately but in parallel as I understand it) and matches colours, tones, relations between shapes, patterns … using a least squares method I think, so providing that a region in the content image is dominated in some way by say a colour then DD will try to find a near match the colour and transfer some the nearby colours, tones, geometry… from a suitable region in the content image using weights of roughly ( a x content + b x style) to produce the result. With kind regards, Andy. Ps. Hi Ben, don’t read to much into analysing the colour and shape relationships, the maths already in the public domain, for instance in the links that I provided in an earlier post). Andy.
Andrew Irving: Hi Daniel and William, Daniel is basically correct in that DD sorts out a place for the colours … but its not always obvious why it choses where to put them. I mostly make composites very quickly (because of time constraints) in which the components have sharp edges – straight, but can be curved or random – because I am aware that DD likes to find edges even in the style images. A good choice of colours, particularly complementary ones, and don’t allow the ‘patterns’ become to large relative to the image size (because they will get washed out) or too small (because they will lose definition and ultimately vanish when too small). ATM I use a very old version of Paint Shop Pro for composite production because it is quick and easy (I also have GIMP and IrfanView and Photoshop and Pop Art Studio (but I let the licence for the latter two lapse a couple of years ago)) …. example of a DDed composite… With kind regards, Andy Irving.
Yonas Tato: For the phone, i use “Pic Collage”, easy to download at appstore. The only small “bug” is that one has to manually snip away their tag, that is like a watermark at the bottom of every collage. Apart from that, it’s free and i would say it is more or less great!