I was on original Voat for a few years before Putt ended it. I though Empress made some really valuable posts and was a good user there. She always had good content. Shame to see her go, but her reason was to focus on her family, her first priority.
Any tips for upscaling? I use controlnet with ultimate sd upscale. I think I tried .06 denoising strength, ESRGAN_4x+ and had decent results, but I haven't experimented enough.
I haven't played around with upscaling too much, but from my experiments I've gotten the best results from just using the builtin "Hires Fix" option on txt2img generation, with any of the Latent upscalers.
It's a bit of a double edged sword though: You tend to get better details, but it can change the image quite drastically. Other methods are better at preserving initial content, but worse at filling in details now that there's more room for them. It's a dice roll whether you like the new details better or not. Anything substantially below .3 denoising adds too much, well, noise, in my experience... but that may be a Latent thing.
I haven't tried controlnet or ultimate sd yet, but both are on my long and ever growing list of "thing to check out".
One thing that worked great on this gen, and you may want to try out, is that second pass with DDIM.
In A1111 terms, the "first pass" was the initial txt2img generation (with hires fix). For the "second pass", hit the "Send to img2img" button below the generated image, then rerun with the same parameters (in img2img) but using DDIM for the sampling method (it's generally the best sampler for img2img changes). I also applied "Restore faces" on the second pass here.
You can also change the prompt and parameters on the second pass, if you like, which will maintain the overall scene of the image, but morph content according to changes. In any case, the second pass adds substantial detail while preserving the scene.
Here's the precursors to the OP, for comparison...
As you can see, it added a lot of details and polish, fixed asymmetries in the face, and applied an "aged photograph" feel to it (which was actually already in the first pass, but so weak and subtle you'd barely notice it).
Sorry for the wall of text. It was shorter in my mind.
This isn't really upscaling as such. It's just generating a new image, using the stretched lowres as the starting point instead of random noise. Maybe that's an academic distinction.
Anyway, it did a pretty good job. Much closer to the original scene, blurry parts remain blurry, focused parts remain sharp, more details added where needed. Much better than anything else I've tried if the goal is to "make it bigger but stay the same".
Of course, we're missing out on the random flowers the previous method added. I liked those for this theme.
Cog was active here some time ago. But either her account was hacked or it was someone who had stolen the name to begin. Started randomly deleting posts in subs they had mod privileges in. Ended up banned, I think.
Empress started watching her niece for like 2 weeks and turned it into a redpill boot camp. Complete with the jew question and dont date black guys. Shit blew up and then she ghosted before the boot camp even finished. Maybe something got back to her irl.
I remember that. A lot of goats gave her a hard time for overdoing it, myself included. Felt bad about being so harsh afterwards... I think she really took everyone turning on her over it to heart. I suspect that's why she went radio silent for a while after. She had good intentions, but dialing it to 11 with a teenager like that is just asking for rebellion.
[ + ] Lost_In_The_Thinking
[ - ] Lost_In_The_Thinking 9 points 2.0 yearsMay 17, 2023 15:51:02 ago (+9/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 4 points 2.0 yearsMay 17, 2023 16:09:46 ago (+4/-0)*
Prompt details:
Negative: lowres, bad anatomy, bad hands, missing fingers, poorly drawn, out of frame, (cropped), ((worst quality)), ((low quality)), jpeg artifacts, text, error, signature, watermark, logo, username, artist name, blurry, out of focus, ((monochrome)), ((grayscale)), ((asian)), ((indian))
First pass:
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3093559660, Size: 640x360, Model hash: 9aba26abdf, Model: deliberate_v2, Denoising strength: 0.4, Hires upscale: 2, Hires upscaler: Latent (bicubic antialiased)
Second pass:
Steps: 20, Sampler: DDIM, CFG scale: 7, Seed: 3093559660, Face restoration: CodeFormer, Size: 1280x720, Model hash: 9aba26abdf, Model: deliberate_v2, Denoising strength: 0.4, Mask blur: 4
[ + ] Lost_In_The_Thinking
[ - ] Lost_In_The_Thinking 4 points 2.0 yearsMay 17, 2023 17:37:19 ago (+4/-0)
[ + ] KeepPoal4fags
[ - ] KeepPoal4fags 2 points 2.0 yearsMay 17, 2023 17:06:20 ago (+2/-0)
[ + ] i_scream_trucks
[ - ] i_scream_trucks 0 points 2.0 yearsMay 18, 2023 01:33:15 ago (+0/-0)
[ + ] Niggly_Puff
[ - ] Niggly_Puff 1 point 2.0 yearsMay 17, 2023 23:15:29 ago (+1/-0)
This resolution works? Nice! TIL
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 1 point 2.0 yearsMay 17, 2023 23:28:48 ago (+1/-0)
You can 2x upscale it like I did here, or 3x for full 1920x1080. 4x would take you to 1440p.
Bonus: You can invert it for smartphone-esque layout. Can produce interesting results, as the AI will make use of the vertical space differently.
[ + ] Niggly_Puff
[ - ] Niggly_Puff 0 points 2.0 yearsMay 17, 2023 23:39:05 ago (+0/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 1 point 2.0 yearsMay 18, 2023 00:15:02 ago (+1/-0)
It's a bit of a double edged sword though: You tend to get better details, but it can change the image quite drastically. Other methods are better at preserving initial content, but worse at filling in details now that there's more room for them. It's a dice roll whether you like the new details better or not. Anything substantially below .3 denoising adds too much, well, noise, in my experience... but that may be a Latent thing.
I haven't tried controlnet or ultimate sd yet, but both are on my long and ever growing list of "thing to check out".
One thing that worked great on this gen, and you may want to try out, is that second pass with DDIM.
In A1111 terms, the "first pass" was the initial txt2img generation (with hires fix). For the "second pass", hit the "Send to img2img" button below the generated image, then rerun with the same parameters (in img2img) but using DDIM for the sampling method (it's generally the best sampler for img2img changes). I also applied "Restore faces" on the second pass here.
You can also change the prompt and parameters on the second pass, if you like, which will maintain the overall scene of the image, but morph content according to changes. In any case, the second pass adds substantial detail while preserving the scene.
Here's the precursors to the OP, for comparison...
Lowres "seed finding" txt2img gen:
https://files.catbox.moe/f6acns.png
https://pomf2.lain.la/f/2mbd4wk0.png
"First pass" txt2img regen of same seed with Hires 2x:
https://files.catbox.moe/fv8ag6.png
https://pomf2.lain.la/f/i00dv4m5.png
And then the OP is the "second pass" result.
As you can see, it added a lot of details and polish, fixed asymmetries in the face, and applied an "aged photograph" feel to it (which was actually already in the first pass, but so weak and subtle you'd barely notice it).
Sorry for the wall of text. It was shorter in my mind.
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 1 point 2.0 yearsMay 18, 2023 01:04:51 ago (+1/-0)
Actually, explaining this gave me a thought: Why even bother with the first pass? Why not let DDIM do all the work in img2img?
Here's the lowres gen from above, fed directly into img2img and "scaled" with DDIM:
https://files.catbox.moe/rai5as.png
https://pomf2.lain.la/f/60edaa2f.png
This isn't really upscaling as such. It's just generating a new image, using the stretched lowres as the starting point instead of random noise. Maybe that's an academic distinction.
Anyway, it did a pretty good job. Much closer to the original scene, blurry parts remain blurry, focused parts remain sharp, more details added where needed. Much better than anything else I've tried if the goal is to "make it bigger but stay the same".
Of course, we're missing out on the random flowers the previous method added. I liked those for this theme.
[ + ] Niggly_Puff
[ - ] Niggly_Puff 0 points 2.0 yearsMay 18, 2023 01:09:56 ago (+0/-0)
[ + ] PeckerwoodPerry
[ - ] PeckerwoodPerry 1 point 2.0 yearsMay 17, 2023 15:55:03 ago (+1/-0)
[ + ] FacelessOne
[ - ] FacelessOne 8 points 2.0 yearsMay 17, 2023 16:08:53 ago (+8/-0)
[ + ] PeckerwoodPerry
[ - ] PeckerwoodPerry 2 points 2.0 yearsMay 17, 2023 17:17:58 ago (+2/-0)
[ + ] Fascinus
[ - ] Fascinus 4 points 2.0 yearsMay 17, 2023 16:09:19 ago (+4/-0)
Miss having her around.
Last I heard from her was a little over a year ago.
She mentioned being active on SearchVoat at the time.
*Edit: Had just posted and was re-reading the comment when I realized @PeckerwoodPerry was probably thinking about KatHarzo.
Thanks to @FacelessOne as her name wasn't readily coming to mind :)
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 3 points 2.0 yearsMay 17, 2023 16:20:06 ago (+3/-0)
[ + ] KeepPoal4fags
[ - ] KeepPoal4fags 0 points 2.0 yearsMay 17, 2023 17:09:40 ago (+0/-0)
[ + ] Lost_In_The_Thinking
[ - ] Lost_In_The_Thinking 4 points 2.0 yearsMay 17, 2023 17:39:11 ago (+4/-0)
[ + ] paul_neri
[ - ] paul_neri 0 points 2.0 yearsMay 18, 2023 02:59:40 ago (+0/-0)
[ + ] MaryXmas
[ - ] MaryXmas 2 points 2.0 yearsMay 17, 2023 17:02:30 ago (+2/-0)
[ + ] Deleted
[ - ] deleted 3 points 2.0 yearsMay 17, 2023 17:11:16 ago (+3/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 0 points 2.0 yearsMay 17, 2023 23:06:06 ago (+0/-0)
[ + ] MaryXmas
[ - ] MaryXmas 0 points 2.0 yearsMay 18, 2023 15:25:50 ago (+0/-0)
[ + ] PhantomXLII
[ - ] PhantomXLII -1 points 2.0 yearsMay 17, 2023 19:52:52 ago (+0/-1)
[ + ] i_scream_trucks
[ - ] i_scream_trucks 0 points 2.0 yearsMay 18, 2023 01:35:06 ago (+0/-0)
[ + ] drstrangegub
[ - ] drstrangegub 1 point 2.0 yearsMay 17, 2023 20:30:01 ago (+1/-0)
[ + ] AngryWhiteKeyboardWarrior
[ - ] AngryWhiteKeyboardWarrior 1 point 2.0 yearsMay 17, 2023 21:28:19 ago (+1/-0)
Still doesn't quite pass as a woman, but send him a pic of your dick anyway
[ + ] Hall_of_Cost
[ - ] Hall_of_Cost 0 points 2.0 yearsMay 17, 2023 21:56:18 ago (+0/-0)
I hope she's doing well.
[ + ] paul_neri
[ - ] paul_neri 0 points 2.0 yearsMay 18, 2023 02:59:53 ago (+0/-0)