- Jan 24, 2020
- 1,401
- 3,794
See how high resolution you can generate with SD and also how fast is the generation etc. What is the limit for this card with SD.
- Jan 4, 2022
- 233
- 869
20min from first txt2img try to final tile upscale. Multiple trials and errors
for testing:
Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 4248416645, Size: 1366x2048, Model hash: 4199bcdd14, Model: revAnimated_v122
Time taken: 1m 5s
Torch active/reserved: 13227/16938 MiB, Sys VRAM: 20536/24564 MiB (83.6%)
No module 'xformers' just stock A1111
Steps: 30, Sampler: Euler a ,1024x1024 txt2img - Time taken: 13.35s
Never use Hi ress. and don't know to set it right
Last edited:
- Jan 24, 2020
- 1,401
- 3,794
I have made many posts about how to use hires fixView attachment 2600558
20min from first txt2img try to final tile upscale. Multiple trials and errors
for testing:
Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 4248416645, Size: 1366x2048, Model hash: 4199bcdd14, Model: revAnimated_v122
Time taken: 1m 5s
Torch active/reserved: 13227/16938 MiB, Sys VRAM: 20536/24564 MiB (83.6%)
No module 'xformers' just stock A1111
View attachment 2600605
Steps: 30, Sampler: Euler a ,1024x1024 txt2img - Time taken: 13.35s
Never use Hi ress. and don't know to set it right
https://f95zone.to/search/269515972/?q=hires+fix&t=post&c[thread]=146036&c[users]=Mr-Fox&o=relevance.
Select "hires fix" of course, use 20-30sample steps,
use 2x hires steps in relation to sample steps (40-60) to keep the composition, use ap 0.3 denoising strength.
The sampler and upscaler makes a huge difference. I use either DPM++SDE Karras or DPM++M2 Karras
and
You must be registered to see the links
, cfg 8-10 but you need to figure this one out for yourself depending on the prompt and concept.Most of the time I use postprocessing GFPGAN and often in combination with face restoration GFPGAN but not always.
Then the resolution. This is where you will find the limit of the card. First use a sample resolution less than 1024,
in most cases it reduces multiple characters and conjoined twins and other monstrosities,
use as high "upscaled by" (hires upscaling) as your card will allow.
When you get cuda error message you have found the limit, reduce it slightly to consistently avoid the error.
These settings gives me consistently very high image quality results.
Of course the prompt is what determines the quality of
the composition and the concept in combination with the cfg scale and denoising strength.
Last edited:
- Jan 4, 2022
- 233
- 869
Hiress.fix take 3m37s and 22GB+ VRAM ,result:
same initial 1024x1024 generated txt2img 13,6s and sent to im2img for ControlNet tile and Ultimate SD upscaling
13,5+53,5 = 1m7s and only 8BG+ VRAM
"old" Hiress.fix take 3m37s and 22GB+ VRAM
- Jan 4, 2022
- 233
- 869
Hiress.fix looks with more clear details when zoom, but Tile looks more shiny overall.
when zoom, but Tile looks more shiny overall.
- Nov 12, 2017
- 407
- 3,182
also would highly recommend a lora called 'adddetail'.. can also be used at a negative weighting to remove details from images (if you ever have a need to do that.)
- Jan 24, 2020
- 1,401
- 3,794
If you want shiny just add "diffused light" or "diffuse glow" or both.. I have not messed with control net yet, it's on my to do list. But other than that, so far hires gives the best image quality with best details imo. It's why I always recommend it and promote it. What makes the big difference is the hires steps. You did not understand me correctly.Hiress.fix looks with more clear details
View attachment 2601018
When using hires, go with 2x the sample steps, meaning if you use 30 sample steps use 60 hires steps. The reasoning is that this keeps the composition. If you don't do this the composition will change from the sample steps to the hires steps.
If you have generation preview you can see it change. It was Devilkkw who gave me this tip I have been meaning to thank him. Thank you Devilkkw!
Also NMKD Superscale is superior to Lanczos imo. It makes a big difference.
Last edited:
- Mar 17, 2021
- 308
- 1,053
If you are interested in photo-realism, i released a positive and negative Textual Inversion for it.
Check it
Edit: aviable for all.
Check it
You must be registered to see the links
.Edit: aviable for all.
Last edited:
Meh, you paid too much for the card. I mean, if they are going to charge that price then give the appropriate amount of value. For $1600 it had better become sentient spontaneously and try to take over the planet.... and here is the place and time to brag for my new 13600k + RTX4090
- Dec 21, 2020
- 1,531
- 3,618
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
Hello everyone! i have some questions about the photos my stable diffusing is producing, why is it producing the collar device around my subject? i only starting messing around with stable Diffusion a couple of days ago. Thanks again!
- Mar 17, 2021
- 308
- 1,053
i guess the shield word in prompt. some type of word like "armor" etc do this.
Last edited:
- Jan 24, 2020
- 1,401
- 3,794
And to add to what devikkw said, it's down to the images the checkpoint have been trained on also. It looks great regardless, good job.
- Mar 17, 2021
- 308
- 1,053
Don't know if is correct place, and if someone need it, i've made a little batch file for fast watermark your image and not strip out metadata. Full detail on
You must be registered to see the links
- Jan 24, 2020
- 1,401
- 3,794
- Aug 24, 2022
- 235
- 614
What's the point?
If you preserve the meta data, someone can just reproduce the image without the watermark..?
Not sure I understand the value / use case.
- Jun 27, 2018
- 2
- 6
anyone have a bestiality lora model to bless me with?
- Mar 17, 2021
- 308
- 1,053
Yes, you are in right, if someone use your same setting, config etc.. maybe reproduce it, but when you add watermark, more data is written inside png, also time of added watermark is written in.
You don't have permission to view the spoiler content.
Log in or register now.