Optimization setting test for getting faster sample result

Playing with stable diffusion WebUI,


3 different version of Power girl Lora

https://civitai.com/models/189428/power-girl-dc-comics-lora


https://civitai.com/models/300420/powergirl-nsfwsfw-dc


https://civitai.com/models/170256/power-girl-cartoon-character-tier-03-patreon-request-or-superhero-or-ownwaifu


prompt :  (full body:1.2) photo of power_girl, girl, solo focus, blonde hair, red (cape), 

cleavage_cutout, (white_leotard), long sleeve, (blue_gloves), red_belt, (blue_boots),

muscular, curvy body, huge breasts, abs, thick thighs, 

__pose__,

{__WAKA/background__|night city, neon lights}, 

(realistic, best quality:1.2), QuickHands, <lora:QuickHandV2:0.6> {<lora:CARTOON_DC_powergirl_ownwaifu:0.9>|<lora:powergirl:0.8>|<lora:powergirl-nvwls-v1:0.9>}


(The weights are adjusted at each checkpoint)




"Power!!" girl



??????






????????????????????????







......??








(Those results are random selecting)

The three Lora that i use, quality has not much differences.


All good quality. (But it looks not for realstic model)




In the middle of playing, 

I was curious to see if I could improve the generation speed faster by steps. (not effected quality)





test setting



check point : pasteldiffusedmix_v22


In 10steps, the main character quality is fine for test sample. 

But the background had so much changed.


Um....




Try again



10steps even batter background than others.


It seems that the steps are not being applied as I intended.




Changing check point


Checkpoint : superAniInvicible_v10


sampler : DPM++ 2M Karras


cfg 7


hires 1.5x


Denoising strength : 0.4


test by steps (This time 15 - 20)



I can't see any consistency in the results as the number of steps 'increases'.




Than how about bigger size?


hires 2.5

(Denoising strength : 0.51)


1.





2.









changing check point model again : MHMix-v1


1.



2.




and etc...



Conclusion : Using a small number of steps to quickly check the results, then select the result to my taste and keeping the seed while increasing hires for step 20 seems to be a wrong approach.


While the quality difference between steps 15 and 20 was not noticeable, the results themselves could change. 


However, it seems fine to keep the same steps when re-generate with increased hires.


An unexpected discovery from today's test is that there seems to be no noticeable difference in quality even with only about 15 steps




Just as the test was about to end, an unfinished image was discovered on 12stpes




Comments

Popular posts from this blog

Solved: Facial Errors in Stable Diffusion Pony Models

Some of results (character full images)

Todays simple discovery (uncannyValley_ilxl10Noob)