Doubts about upscale and advice

[barrubba]

I am creating graphics for printing, starting rectangular format 1024 x 1280 with Highres fix activated.
I have a few questions for those more experienced than me:

I have a 3060ti 12gb, how far could I go with the resolution?
Sometimes I notice watermarks, what should I do?
Even if I add "sharp focused" in the prompt, the edges of objects are often not very defined. I also use (((blurry))) in the negative prompt.
Also, if I want to upscale twice to print on large formats, do you recommend using a denoising value close to 0 so that strange objects are "born"? The basic scene as composition and objects suits me very well I don't need to add new elements but only to define edges, shadings and more micro-detail, but not "invented". That is, if there is a small/medium sized fish, I would just like it to become more defined and with more detail, not to see a face or something unrelated appear on its side.
For now I'm getting on really well with the 1.5+ Dpm2aK, HiresFix 0.0, 0.2
Upscaler I generally always use 512/32 R-ESRGAN 4x+ Anime6B.
Do you have any advice on this!
SD has disrupted my creative and imaginative canons!
Wonderfully Disrupive!

[gregwbrooks]

So, I'm you -- printing stuff traditionally, often at large sizes. Here's what I do:

1. Generate core image -- if I'm willing to wait (and depending on the composition) can can get about 60-80k pixels of original image.

2. Run it through img2img using the SD Upscale script and a Remarci upscaler. (I'm doing stuff that looks painterly, so that's the best fit. Depending on your subject matter, you may want a different upscaler.) I vary the settings, but typically it looks like:

• Text prompt: "highly detailed" (no quotes)
• Euler (not Euler A)
• 150 steps
• CFG of 11.5 and denoising strength of 0.15

3. Cycle the upscaled image through the same process a couple of times. You're adding detail each time, but as long as you keep the CFG high it's not like a new fish (to use your example) will pop up. It gets time-consuming, (going from, say, a 4096x4096 image to an 8192x8192, for example, might eat 2-3 hours of your life). But you can get great results.

4. Where you stop depends on what you're printing, how much detail you want to create, etc. I generally use this process on my best stuff to get into the 6k-to-8k-pixels-on-a-side territory, and use Upscayler it I need to go further.

[ClashSAN]

one thing to note, a bug doesn't let you use other samplers for SD upscale's img2img. Your default I'm guessing will be Euler A. Rollback (1wk 828438b) (2wk 98947d1)
Before SD upscale was not a script, you should consider locating the commit version. Why? It lets you input any sized image, skipping the first 2x upscale process. You'll be able to enhance detail of pictures of unlimited sizes. You can still do this by setting your choice upscaler to use cpu, (to avoid gpu OOM) but it will take lots of time.

[barrubba]

Please can you explain better how to use this "infinite" upscaler? I've not updated a1111 version for 1 month..

Replying to gregwbrooks:
Thank you!!!
In my case I realize digital paintings highly realistic, not pictorial. So I will try remacri (unknown to me until now, thanks ).
I don't understand how you can get 80k pixels on the first generation step. What video card do you have?
I use - - xformers to increase performance, if you don't already, definitely try it.

[gregwbrooks]

@barrubba, apologies -- I typo'd the pixel number and it probably just made it confusing. I should have said 600k-800k pixels. For example, if you generate at 832x832, that's 692,224 pixels; a 512x512 image would be 262,144 pixels.

[barrubba]

So, I'm you -- printing stuff traditionally, often at large sizes. Here's what I do:

1. **Generate core image** -- if I'm willing to wait (and depending on the composition) can can get about 60-80k pixels of original image.

2. **Run it through img2img using the SD Upscale script and a Remarci upscaler.** (I'm doing stuff that looks painterly, so that's the best fit. Depending on your subject matter, you may want a different upscaler.) I vary the settings, but typically it looks like:


* Text prompt: "highly detailed" (no quotes)

* Euler (not Euler A)

* 150 steps

* CFG of 11.5 and denoising strength of 0.15


3. **Cycle the upscaled image through the same process a couple of times.** You're adding detail each time, but as long as you keep the CFG high it's not like a new fish (to use your example) will pop up. It gets time-consuming, (going from, say, a 4096x4096 image to an 8192x8192, for example, might eat 2-3 hours of your life). But you can get great results.

4. **Where you stop depends on what you're printing, how much detail you want to create, etc.** I generally use this process on my best stuff to get into the 6k-to-8k-pixels-on-a-side territory, and use Upscayler it I need to go further.

Why not Euler?

[gregwbrooks]

The Ancestral samplers like Euler A can add a bunch of noise per step, so they are more chaotic and diverge heavily from non-Ancestral samplers in terms of the output images.. For img2img upscaling, that divergence from the source isn't usually desirable.