[ 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.

/vt/ - Virtual Youtubers

Search:


View post   

>> No.43679416 [View]
File: 1.20 MB, 1377x4000, xy_grid-0283-4161217259-high quality, best quality, high detail, very detailed, official art, ceres fauna, 1girl, green hair, antlers, yellow eyes.jpg [View same] [iqdb] [saucenao] [google]
43679416

>>43679317

>> No.40202043 [View]
File: 1.20 MB, 1377x4000, xy_grid-0283-4161217259-high quality, best quality, high detail, very detailed, official art, ceres fauna, 1girl, green hair, antlers, yellow eyes.jpg [View same] [iqdb] [saucenao] [google]
40202043

>>40201489
>>40201850
>images for ants.jpg
I dont think its fair to compare a 4000x max resolution jpg of upscaled images

Some of my takeaways from this.
Most upscalers didnt deviate from the average generation time of 2 seconds firstpass and 16 seconds finalpass.
1. The Latent upscales generates more detail and I think this is the one I will keep on using
2. Nearest sucks
3. LDSR loads a seperate model and does a 100 step DIMM pass for an extra 33~ seconds.
4. SwinIR also does an extra pass for 6 seconds
5. Higher denoise strenght adds in extra detail in the background but also fucks hands more.
Any other differences could be xformers playing a trick on me but Its not worth sacrificing the extra speed for determinstic generation. Thats why I atleast did the tests on Euler.

Navigation
View posts[+24][+48][+96]