the documentation says Q is more general, F better detail and avoids degradation (hallucination)
the documentation says Q is more general, F better detail and avoids degradation (hallucination)
Pyeah i think f gives more detailes
RI find that this applies mildly
Pwaht checkpoint you are using?
RI find that x1 upscale is prone to artifacts like monstruous teeth and overkill eyelashes on lower quality images in midshot to fullbody
Rso x2 is almost completely necessary if you're running a big batch with lots of small faces
Rjuggerv9
KF is for fidelity. Better at staying closer to the original image details
Kyou should use default sdxl model. Juggernaught tends to create worse faces
Rcaption makes a considerable difference, so run @Dr. Furkan Gözükara 's Kosmos app beforehand and add to the image input folder
Fstill trying to fix errors haha
RI have seen very high fidelity on faces
Ri've kept a close eye and have not seen much degradation of appearance
Rdo you use mostly F model?
RSo no surprises there on any exception cases?
RHave you played with cfg?
KI mainly use the quality model with kosmos. Seems more predictible and better results overall.
cfg i use default at 7.5 and steps i use 40.
But workflow for full body images. I use photoshop and split the image in 2
cfg i use default at 7.5 and steps i use 40.
But workflow for full body images. I use photoshop and split the image in 2
ROriginal WebImage vs Fmodel x2 Upscale
Rand Q model
RRunning on Juggernautv9
RHard to tell the difference
RCould not see much difference in using stage1 + stage2 over just stage2. Anyone get better results?
KIf your images is poor quality you need to run stage 1
Kif its a decent images just stage 2
Kbut overall running both 1 and 2 is the best
Rnot necessarily a bad image this one. What do you think?
Pso for sd 1344x768 i can just run stage 2?
Khow much you upscaling by?
KFull body much harder
Rtesting stage1+stag2 now
Ki'd split that image in 2
Pso far tested only 2x
Rcant do that, have a 600 image dataset
KYou need to examine your image if its already decent quality, then yeah stage 2 only
Rbut to be fair x2 is quite excellent result, its just very time costly. Takes 125 seconds. x600 takes 1.5 days on an a6000
R6 hours if only x1 upscale
Rat 79 cents per hour, its the differnce between 6 dollars and 30
KOkay wait lemme use the comare site
K
Ktook me 80 secs
KCan i try out your image?
Rsure. i pasted it up in the chat
F
Fdone
Pgigapixel vs supir 2x from 1344x768 https://imgsli.com/MjQ0MTY3
F
F
Kstill working on the stage 1 for batch?
Pis 3x or 4x possible with 24 gb vram? (from 1344x768)
Kmax i go is 2.5x
Fi am going to add it soon hopefully working