[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/sci/ - Science & Math

View post   
File: 1.64 MB, 1316x1152, craiyon.png [View same] [iqdb] [saucenao] [google]
14732710 No.14732710 [Reply] [Original]

why is the 0.4 billion parameter craiyon able to do a fairly decent job? Note the parti-images were built up using some sort of in-paint feature whereas craiyon was all in one prompt.

>> No.14732771

>>14732710
Google engies are probably retarded or are doing a purely scaling approach.
Craiyon might be using heuristics on top of scaling.

>> No.14732788

>>14732710
Now ask it to do human faces, or use negation statements.

>> No.14732915
File: 688 KB, 598x608, craiyon2.png [View same] [iqdb] [saucenao] [google]
14732915

>>14732788
not bad actually 5/10 for Craiyon.

>> No.14732919

>>14732915
That banana melting into the board is so dalliesque

>> No.14733012

>>14732915
I mean 5/9
>>14732919
Yeah those are interesting to me. I recall doing something similar when I was a 3. When I would get anything wrong I'd draw over it.
If you look the bowl is yellow so somewhere in the steps to create it the banana ended up there and not there at the same time.

>> No.14733030

>>14732710
Parameter count is not a useful heuristic for comparing models with different architecture. Given a particular network design, scaling parameter count seems to pretty reliably improve performance.