Make Bigger Images in SD: Pros and Cons for Outpainting, Hires Fix, Img2Img, and ControlNet Tile 

Something to add, the multidiffusion-upscaler-for-automatic1111 extension under Available in A1111 comes with something called Tiled VAE which with default settings allows my 8GB GPU to do 2x Hires. fix upscaling where it normally would get CUDA memory errors. Worth taking a look at as well, and I assume this can allow a high end GPU to push even further as well. This option is also available under Img2Img, though I have not experimented with it yet there. My assumption would be that it allows for even higher upscaling in that tab also. Multidiffusion is another tiled upscaler solution btw., although I haven't played too much with that part of it yet so no idea how good that is compared to the Ultimate SD script.

this is spectacular information, thank you for going into 10x more depth than anyone else I've seen along with visual examples and guides.

I love your diagrams and explanations and humor. A+

Seriously awesome content man. Keep it up. Not enough good video tutorials for A1111 and SD out there. Subbed.

Note that you can add upscaler_for_img2img to your Quicksettings list so it will show up at the top instead of having to go to options to change img2img upscaler. (i usually add vae and clip skip too)

There is an incredible amount of info in this video - all very clearly laid out with out any "Puff & Fluff" like alot of other ones. I learnt a lot of snippets of info. I find another handy feature of the image to image tab for resizing is the Batch tab. I'll generate a large batch of images at a lower res & skim through them and delete a heap that i dont want, then I'll "Batch Upscale" the whole folder using the Img2Img features (the upgrade to the slider for size was a saviour for this because it means you can mix Portrait & landscape orientation in the folder & you can also do a bit of cropping when you are choosing the generations to upscale & it will just do a blanket multiplier regardless of the orig dimensions). Keep up the good work 🙂

man, i always had the trouble of not having script under controlnet like in all the videos i saw... but its literally just at the bottom, i just had other stuff before. I stupid, but glad i found it.

details and helpful info, thanks.

This was awesome. Simply the best workflow for great consistent upscaling.

Amazing the overview bro 🙏🙏

Thank you so much for all the effort that you put into this!

1) I get best results with latent upscaler - HiresFix settings: Latent (nearest-exact), denoise 0.63, hires steps = txt2img steps / 2 You can use Latent upscaler in img2img too. 2) The choice of sampler affects the upscale result: try different, in that order: DDIM / SDE / 2M, 2S Euler a sampler makes everything more blurry, imho 3) Use Tiled VAE (adjust tile size) - if you have CUDA memory error 4) Use different checkpoint (model) for upscale in img2img, for experiments

Hey man, earned my subscribtion. One of most elaborate ordered and thoughtful essays on the theme.

@SiliconThaumaturgy, I started learning about SD--and AI in general, just some weeks ago (literally from square zero). You were among the very first RU-vidrs I saw. I'm very technical & thorough, your videos were technical & thorough. But more than that, it was practical. There are many AI channels out there just as technical as yours, but the vast majority of what they're talking about can't really be applied to what I needed to do. There are also many who's just as practical as you, but they lack your technical depth. So in that sense, you're unique, or at least very rare. When I wanna learn something, I do it really fast, w/ zero BS. So I sat down and finished every single one of your content in one sitting (yes, every single one), after eating a quick midnight snack. (My friends say I'm scary like that.) But I'm still subscribed to you, and besides your newer videos like this one, I still sometimes go back to your earlier ones for stats/reference. Like your Samplers video, it's getting a bit long in the tooth, but it's still useful when I need to review which "group" a sampler belongs to (I know it's not really an official grouping, but that's how you termed it). So keep your content like that, and you'll keep subscribers like me--to infinity & beyond. Now a query, if I may: I'm actually glad that you did an Upscaling topic, because there's not much content out there regarding it--at least NOT the technical and thorough and practical kind that I like--and I was actually hoping you'd do one. So it's really great that you did, but my question is, why didn't you include MultiDiffusion Upscaler (MDU from this point on) in this analysis/guide video? I'm presuming you already know it (coz I'm really new & I do), but just in case you don't, it's a suite/package of VERY useful extensions for SD/A1111, like Tiled Diffusion (this is the main upscaler), Tiled VAE, Regional Prompt Control, etc. I know you said in the video that you're only including "default" upscalers/extensions like Ultimate SD Upscaler (USU) & ControlNet Tile (CNT), but what disqualifies MDU from that criterion? 1) You still need to separately install USU & CNT (not to mention their associated models), just LIKE MDU. In fact, beyond that, if you go to A1111's Wiki section, MDU is actually listed there as a major Feature of Automatic1111, just like USU & CNT, so it's not any "less default" than them. And that's putting aside the fact that its ubiquity AND popularity in many install bases make it "de facto default". 2) Just like USU, MDU teams up really well with CNT. In fact, I've read creator/user reports from CivitAi, Reddit, etc. that it's actually better than USU. I have no scientific data to back that up (I was actually hoping you'd do an analysis on 'em weeks ago), but I HAVE used all 3 of them, and at the VERY least, I CAN attest to the fact that MDU can go toe-to-toe with USU, in terms of quality and/or performance. 3) If you want to, MDU can be used very early in the workflow, even in T2I (with or without CNT), UNLIKE USU--for what it's worth. 4) As I've mentioned, it's actually a suite of tools, and its siblings ain't no slouches either (check 'em out). So I'm a BIT disappointed, because I really wanted to hear your thoughts on MDU, ESPECIALLY in relation to CNT & USU. Obviously, you can't modify the published Upscaling video now, but perhaps you can do an addendum video, or something like that. Anyway, aside from that single strange omission, this video is still very technical and thorough AND practical, and I still really like it. I'm definitely bookmarking it for later reference. So good job. BTW, while we're on the subject of Upscaling, you once said in one of your EARLIEST videos--I think it was the Hires Fix one--that you were gonna put together something about Upscalers, as "they deserve a whole other video of their own", or somethin' to that effect. In fact, you even kinda alluded to it again in THIS video right here. (Yes, I never forget.) So, consider this the text(ual inversion) equivalent of that old GIF meme with the old guy, "WELL? We're waiting..." Heh. Keep up the awesome work, have a great day, and cheers. 😊 PS: You might also wanna update/addendum your older CFG Scale video w/ Dynamic Thresholding, and your newer Inpainting video w/ LLuL (Local Latent upscaLer) & ADetailer (After Detailer). Yes, I've used them, but I'm still very new & learning (just like a new AI), and I like and respect your analysis & advice. So I'm still very curious about your final judgement or recommendations regarding them.

this is what I'm looking for! I'm learning this A111...thank you for this..

Thanx for this amazing video

Great video, thanks!

WOW! It seems you covered everything! This is a great video and might be the best on youtube for comparisons on this topic! I have tried most explained in this video. I work alot with portraits and I gravitate to img2img upscale x2 with denoise around 0.1 to later upscale again with UltraSharp x2 in both steps using restore face. ControlNet is super important for good and consistent output. The image output are amazing and the faces are almost so good they become 3D popping out of the screen. However I'm working infront a 43" 4k TV screen. Anyway great explained!

Love this type of content!!

Very interesting video. I do find that Latent upscale gives by far the best results for photoreal images, with a max scale factor of 1.5 and denoise of around 0.45-0.5. The issue is it tends to ruin a decent % of images. Do you know if there is any way to use tile_resample inside Hires Fix?

Thank's alot

you're my hero, I'd do anything to get on a call with you for 15 minutes