bro me got question why more vram why bigger better?

[mypromise88]

sorry quetion is stupid gonna make it more clear

so the thing is, i see bigger vram is important in ai image generation,

what they say is, bigger vram can generate bigger image with bigger resolution like 1k x 1k and even more,

fine, i get it.

but what's the difference between 512x512 and say 1k x 1k or bigger image?

i mean i know the default size is 512x512 and you just get 512x512 and can enlarge the image without quality loss.

but just wondering why everybody seems crazy about having big images like 1k x 1k or 2 k x 2k and so on,

i am noob in this, so enlighten me on this if i am missing anything bro

[ClashSAN]

A larger txt2img will often come out sharper by default. However, take note, the subject may be skewed. If you generate a higher resolution 1024x1024, you may experience an unwanted cloning effect, unless you do things like seed resizing or highres fix, features found in the webui. The trend and enthusiasm for larger images may be:
-people enjoy performance testing or capacity tests, The --xformers and --lowvram will allow even larger sizes
-assurance your gpu can handle different sizes below the range, like 512x768 portrait sizes, or 720p sized wallpapers, panoramas

-people want more detail in their pictures. for img2img, enhancing a generated image at that 1024x1024 size will add "creative differences" (you don't actually need lots of vram, instead you can img2img in chunks with SD upscale script)
(example: make a dot on the wall a pattern at higher resolution. a standard upscaler will open just make it a bigger dot.)

There's a buttload of things to talk about regarding the subject. got more questions? ask away

[yspacedev]

I'm currently using an RTX 3060 Mobile GPU with 6GB VRAM, and it works just fine. It's probably not going to be able to train SD, but it works extremely well with generating images. I'm getting around 4-5 its/s on Euler a and similar fast samplers, and half that on slower samplers like Heun. The desktop version of the RTX 3060 has 12GB VRAM (more than enough to do all the things). I don't think there's any reason to shell out the money for a 4090 or 3090 just to generate images. An older generation GPU will probably serve you just fine because SD does not use the RT cores or any of the newer technologies in GPUs.

[hvanmegen]

A 4090ti will never have 48 GB of VRAM since more than 24GB is reserved for Nvidia's professional lineup of cards which cost over 5000 dollars.. they're not going to compete with themselves on that.

[ClashSAN]

@hvanmegen that true. did you figure out that user-config.json to set the max sampling steps ?

@0xalivebeef RTX 3060 good recommendation. It will do the 3 style training features that @mypromise88 asked, for a cheaper price.

[yspacedev]

Another recommendation might be the RX6600. It's cheaper than the RTX 3060, but has about the same rasterization (and therefore compute) performance. It just doesn't have the ray tracing power and support that the RTX 3060 has. But there is an install guide for AMD, so an AMD GPU should work.

[mypromise88]

THANKS BRO
I WILL ASK MORE QUESTIONS SHOULD THEM ARISE

[78Alpha]

A oneshot comparison would be

512x512 (no Hires.fix)

1024x1024 (HiRes.fix applied to keep them the "same")

[mypromise88]

thanks now i see why bigger vram god damt it!