This A1111 Trick is NEXT LEVEL

[@OlivioSarikas]

#### Links from my Video #### Download 8x_NMKD-Superscale_150000_G.pth from here ( use the arrow pointing down on the right of the model name. Download it into the A1111 - Model - ESRGAN Folder) huggingface.co/uwg/upscaler/tree/main/ESRGAN My Prompt: docs.google.com/document/d/1VfOtwTPyPHTGBe3kbc1utwVZLemmYSk-3bWHgm6WsIo/edit?usp=sharing EpicRealism Model: civitai.com/models/25694/epicrealism

[@LouisGedo]

👋

[@mercuryman1250]

What program is this he's talking about?

[@AnimeGamerHub]

is this better than your previous 4x ultrasharp?

[@MiguelSilva-vq6qe]

Great !!!

[@KlausMingo]

How do you install the upscaler file?

[@idanis5948]

I think your hair affected my algorithm because now I'm getting recommend electric callboy music videos

[@OlivioSarikas]

😂😂😂😂😂😂 Awesome!

[@lurker668]

😂😂😂

[@xXxDisplayNamexXx]

Bro I’ve been into them since 2010 and I’ve loved them ever since. They’re finally getting big now and it makes me so happy 🙌🏼

[@henrischomäcker]

If your favorite photorealistic model does not listen well to your prompt, use the revAnimated or even disneyPixarCartoon model to create your desired scene with txt2img and then send it to img2img, change the model to your favorite photorealistic model and then go with a cfg scale of 7 - 10 and a denoising strength of 3 - 5. You may even directly increase the size with this step to for example 1024x1536. That even works with minimum vram. If you choose a higher denoising strength of around 5 you may also add the addDetail or moreDetails Lora to the prompt while doing the img2img step.

[@jakesalmon4982]

So weird that you commented this, this video gave me this exact idea earlier today. I used revAnimated to create my fantasy RPG character and sent it to img2img with a realism model. Worked really well with some tweaking.

[@TheDocPixel]

I’ve been doing this with my own artwork from Photoshop and basic sketches using txt2img with ControlNet first. Then sending the result through all of the checkpoints using XYZ script to choose the best ones and returning to Photoshop for touch ups. The workflow makes it copyrightable as an added bonus.

[@paskmoe4969]

I will give that a try thx

[@MonsterTracksKaraoke]

You can even do this using hi-res fix, since the checkpoint and even prompts can be different.

[@ZiggySA]

Bro just use the refiner and switch at 0.5 to the realism model you want. No need for img2img.

[@ens_lucis]

I'm already blown away with all that's energy you have put in this video. IM LIKE SUPER CHARGED WITH POSITIVE EMOTIONS RIGHT NOW. The way you tell about this is really amazing, you are THE BEST!🔥

[@OlivioSarikas]

Thank you so much 😍

[@Sheevlord]

It also works when you generate an image using one SD model and then put it through img2img using another SD model. Some models have very creative compositions but fail at small details, so this approach can help a lot.

[@nalisten]

Olivio!!! Your videos hold so much Value🙏🏽🙏🏽🙏🏽 I appreciate you. I been working on an upscale method you posted months back to enhance the quality of all my past work. even with all the daily tricks coming out in other videos, nothing is as powerful as your workflow in my opinion. Much respect✊🏾💪🏾👑

[@tomaskrejzek9122]

That's exactly how I love you, Olivio. Funny, over the top, full of enthusiasm and energy. You're the man!

[@mazireth]

Didn't watch the video but thought you deserved a like for that hilarious wig. I got a good laugh

[@PallaviChauhan91]

You are the best thing I watched today :D Simply awesome!

[@juanjesusligero391]

I'm a simple man. I see Hackerman, I press like button.

[@BlueSpork]

I love your new personality much better than the old one ... not saying that there was anything wrong with the old one :) Thank you for this video!

[@OlivioSarikas]

thank you 😅

[@polloloco6353]

I've been doing a lot of testing based on your video, different denoising adjusters, different models including DSXL base 1 (which leaves the original anime image almost the same and doesn't turn it into a real image at all), adding more negative prompts, and I can say with certainty... THAT YOUR FIT IS SIMPLY PERFECT...! Thank you very, very much Olivio...!

[@polloloco6353]

But for upscaling, I prefer Ai Photo & Art Enhacer standalone software...

[@tharidan]

@@polloloco6353 is it better than ultimate sd upscale ? is it low vram friendly ?

[@polloloco6353]

@@tharidan This was a year ago, I don't remember... sorry.

[@jonescrusher1]

Thanks for this Olivio, by far the best process i've tried so far.

[@AIKnowledge2Go]

Great video, Olivo! As always, I appreciate your unique style and your 'Hackerman' outfit commitment 😊. For finishing touches, I prefer the Tile Upscale method with ControlNet over your 'send to extra' method. Sending your image over to img2img, as you demonstrated, really should be the default workflow in A1111, regardless of the image's origin. When I'm satisfied with the overall composition but want to enhance details, I set the denoising strength between 0.25 - 0.35 and resize by a factor of 2. (This can vary based on the number of Loras you use; more Loras with higher denoising strength usually results in more deviation from the original). Many folks opt to use the high-res fix in txt2img, but in my opinion, it doesn't deliver the same quality and is a less efficient workflow.

[@Mp3Pintyo]

How do you achieve that the line between the tiles is not visible? With a homogeneous surface, you can see the joints in the case of ControlNet Tile.

[@AIKnowledge2Go]

@@Mp3Pintyo i don't wan't to advertise my channel here, but i actually have video about it where i encounter this problem and for me it had an easy fix. Its called: Best Practice Workflow Part II - How To Upscale Your AI-Art.

[@ChristerHolmMusic]

Nice tip. Working beautifully! It's quite crazy, the feeling is like you are getting the original photograph, from which the Midjourney art was painted :)

[@vs3183]

It would be an understatment that I wait for your videos just to see you perform in this persona. Pure joy! Thank you : )

[@jcdeton72]

You look great today! :D Thank you for the video and keep up the good work!

[@LeKhang98]

Damn I like your hairstyle and your "mwah" flying kiss lol. Very nice trick thank you for sharing.

[@GuyTheAnimated]

you've taken me to the next level I might have to try with source images from the Wurstchen demo

[@ArthurDNB]

The wig was amazing hahahaha

[@OlivioSarikas]

Thank you 🥰

[@kewa_design]

Still can't believe it, feels like yesterday when I was totally amazed by first DALL-E stuff, which is baby drawn pixelated, ugly mess today. And it's like a year ago or so 😂 now you have this shit and it's incredible. I don't know how much better image generation will get, but it will faster in the future. Anyways another incredible, video like always thank you ❤

[@Vestu]

This is amazing and something I've been thinking about as a long time MJ-subscriber and quite fresh SD-explorer. Please be this this personality with the wig in every future video 😄

[@JoaoPauloDev]

Works for me using forge, it is really mind blowing 😉 Thank you, I was really looking for that!!!

[@mikesrandomchannel]

I love the idea of MJ being schooled by Stable Diffusion 😜. I've also been almost more impressed by the upscaling in SD than the image creation, it's incredibly useful in so many ways. Great video, always love learning more!

[@matthallett4126]

If you run out of VRAM use the script extension Ultimate Upscaler, and Controlnet TILE.

[@quercus3290]

tile VAE would use less no?

[@matthallett4126]

Perhaps, its kind of the same thing? I use the tile controlnet on larger images and upscale in 768 chunks. That way you can go from 2K to 6K. @@quercus3290

[@60FpsGoodness]

@@quercus3290 You can also use Tiled VAE. The Controlnet TILE is almost mandatory when you use something like Ultimate Upscale.

[@tac-22]

Olivio your style is unmatched! ♥🤣 Perfecto.

[@AI-Stories-Fun]

My god it blew my mind. I upscale on daily basis and definitely need to try this. I'm not able to follow and try everything about AIs. It's just too much incoming too fast 😂

[@danilolopez9203]

Really good trick, thx 😊

[@DJHUNTERELDEBASTADOR]

conversion?? 😅

[@thewebstylist]

I hope MidJourney adds something like this for an all in one! Great vid

[@storeyhome]

Would love to see a comparison between this and Leonardo AI's new Universal Upscaler. PS: Nice hair!

[@BeniBond]

I totally can feel with you.❤ You are magic my creator friend 😊

[@RajeshJustaguy]

i tried this and it's pretty impressive. and very fast and quick to do..

[@huminbyrd67]

Top content creator. Cheers.

[@MarcoEscobedo10]

I did the same but with the images generated with SDXL 0.9 to improve them

[@ImpoMundus]

that was fun - I like the surprise!

[@gasia112]

With resize by x1.5 (or higher) + controlnet tile (weight0.6-0.7) and img denoising str 0.55. It will add more detail to your img too

[@TheKamrenB]

How about mixing that with Ultimate SD Upscaler?

[@Mp3Pintyo]

Can't you see the joints between the tiles? For me, they are almost always there when the surface is homogeneous.

[@gasia112]

@@Mp3Pintyo that's because you use ultimate sd upscale, right?. change mask blur to 8-12 (but i didn't use this method)

[@Xinkgs]

😅 LOVE THAT HAIR AND THE CONTENT OF COURSE ❤

[@memoryhero]

Okay - Hackerman's gotta stay. I...need him in my life.

[@0AThijs]

I use refiners for this, get a base model then use the refiner model to change its style. This works way better of course, thanks for sharing!

[@swannschilling474]

Will try it later!! Looks great!! 🎉

[@RegisHervagault]

Hats off !! this is why we love you

[@comiccultivation]

Using the img2img tab is a trick now?

[@VulgarTruth]

Hahaha youre a funny lad! Subscribed

[@SB-qm5wg]

That is pretty cool

[@ai-pix]

Oh, wow! Quite an impressive result!

[@Naturalezaenaltura]

Awesome results!, Thanks as always for your tutorials!!!

[@OlivioSarikas]

My pleasure!

[@giusparsifal]

Hello and thanks, when I send to extras and apply the upscaler nothing happens, it generates the same image in 0.0 second. Any suggestions? Thanks

[@Svior]

Bro 😭 i really love these type of videos man! Sometimes when the other guys make a new character out of themselves it's somewhat cringe and just a little too forced. But you just nailed it with the energy! Big love Olivio ❤

[@tharidan]

so if I generate an image from a checkpoint, can I use this method? Should we keep the originally generated image prompt or write a short prompt like in the video?

[@xcrap]

You can have similar or sometimes even better results using txt to img and using midjourney or other photo in control net with softedge and even using multiple for a more precise control.

[@middleman-theory]

Do you mean multiple softedge images? Are they duplicates of the same image, or two different images? Please explain further...

[@xcrap]

No, just one image. So like img2img, use the reference image as a control net with softedge type. If that softedge does not have enough definition (some images may lack edges) you can add another canny control net (with always the same reference image) or sometimes I use open pose with dw_openpose_full for correct hands and body position. It will give not exact results like the img2img regarding background but I find to get wider and super interesting results. I do this with normal photos as well. Using the base prompt "RAW photo, SUBJECT, 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT3" you can even use the img2img interrogate clip to give you the info for the SUBJECT area. He should make a comparing tutorial with the two techniques using the same reference images, it gives very interesting results. My english is not the best so.... that's it :)

[@ardezart]

Olivio, thank you for your work! All your video is awesome! Do you have any plans to make a new outpainting video in SD, similar to photoshop techniques without the prompt to quik clean the image from artifacts and increase the size of the image?

[@MarylandDevin]

What was the trick supposed to be, Image to image, upscale, the model? I'm pretty new to this, but nothing stood out to me here.

[@toddg6575]

Hi Olivio - What would be the best way to do the reverse of this? Which model and prompt do you think would work best to take an ordinary photo of someone and turn it in to a magical creation? Thanks.

[@palax73]

Amazing. Thank you so much bro.

[@gkotsetube]

Any help on how to incorporate the downloaded upscaler in Invoke AI?

[@Steamrick]

If you like high detail, upscale using img2img with higher than you'd otherwise dare and the tile controlnet. Compared to any upscaler, this results in far more detailed images. If you run out of vram due to resolution, use tiled VAE and/or the ultimate SD upscale script. Tile controlnet helps with those, too. edit: And throw in the more detail loras for good measure if that's what you're into.

[@CHARIOTangler]

Hackerman rules the AI space-time continuum🤘🤘🤘

[@LyoshaTheZebra]

The upscaler is not visible here in the UI, what might be the case? I have put the model into the models/RealESRGAN. Still, it just doesn't get listed.

[@OlivioSarikas]

put it into the ESRGAN folder and then restart A1111

[@jibcot8541]

I would probably add a 2x resize on the img2img also, but apart from that it's good, I have been using this technique for quite a while for blending between different models and adding detail.

[@romanlevytskyj8625]

Damn my finger tips are itching to test it on some old generated art. 🎉

[@ThoughtFission]

Could you do all of that in one worflow in Comfy?

[@MrVitaliyAT]

Cool idea cosplay somebody ))) looking easy and relaxed ))

[@thomas_k.6862]

Sorry, it doesn´t work. I´m missing the models, Olivio was talking about. They are not installed. Has anyone a link where I can download them manually or any other helpful advice? Thanks!!

[@botnugget3784]

I wish they would make an extension to automatically redo the render in a 2nd checkpoint after the first one. Refiner switching is good but it would be so much better to have a full render in the 1st checkpoint and then get an img to img 2nd checkpoint render automatically.

[@Atlantis_AI]

Thx i will try it, epic model work very well on RunwayML videos too

[@TheColonelJJ]

The forever new question: Where would this go in ForgeUI?

[@johnwilson7680]

What VAE are you using? Auto? Thanks.

[@etherhealingvibes]

Just like in economics, imagine a scenario: "When everyone can obtain gold, what is the cost of gold? Zero." If you can create art without being an artist and make music without being a musician, what is the cost of art and music? Zero. We are all heading straight towards a massive iceberg, and we, as human beings, are on board the Titanic. Undoubtedly.

[@interactiveGTS]

Personally I thought it went to hell when we started letting the most evil men run our nations and conduct endless wars for profit killing millions of people worldwide, but Im sure some free pics will really destabilize things too.

[@gaelpiratelife1941]

Nice i idea, i will give it a try😊 , ( you have too much tabs in your browser 😅 )

[@micbab-vg2mu]

Fantastic!!!

[@Gamepunk66]

Now my secret trick is on KZbin

[@Elwaves2925]

Nice trick, thanks for highlighting it.

[@Pilek01]

So im new to A1111 and followed all this steps and it takes like 40 minutes or more for 1 image. is that normal? i have a 5800x3d and 7900 xt, 32 gb ram.

[@Ахриманстоппер]

can you order to ai to make exact face model and other transformations on few pictures? kind of can you render a movie or cartoon and watch it in different style, cast etc?

[@styrke9272]

Great!! With this is possible to bring more details to photos from a old phone?

[@alejmc]

Old phone pictures you mean? Just the upscaler part should be enough from what I understand… Topaz software already was doing fancy upscaling out of the box years before this AI craze exploded. Short answer: surely

[@styrke9272]

@@alejmc thanks for the reply

[@quercus3290]

@@styrke9272 Personaly I would attempt various upscales, experimenting with both de-noise and CFG scales until you find an acceptable image. Then once you have the result you are after, follow the steps in this video with your recently upscaled image.

[@OnigoroshiZero]

Thank you for your work, it's appreciated. Btw, how long does it take you to do an img2img generation at that resolution? I have been out of the loop for 40-50 days, because I was annoyed by the increase in generation time that HiresFix and img2img at higher resolutions (or x2 upscale) would make. Using 30 steps with face-restoration plus 1-2 ControlNet options, it would take 4-6 seconds regardless of sampler at 512x768 on my 3080 10gb (using RevAnimated model). With HiresFix or an img2img pass (with x2 upscale) this time would increase to anywhere from 2 up to 15 minutes, and this increase was completely random (just with a different seed it could have 5-9 minutes difference). I was using x4-UltraSharp for everything regarding upscales which is supposed to be fast from what I remember (found it from another of your videos when it first came out). Any comprehensive guides using the latest models and settings for such things would be nice. Also if there is any new models similar to RevAnimated but for XL that has better results (mostly hands and similar features in which Rev struggled a lot).

[@xhavierjanssen6305]

GREAT !

[@mkaleborn]

Great tip! I really need to try doing that with a few of my illustrated/painted/digital creations. I can see SDXL being a good starting point as well, with it's more creative pseudo 'Midjourney' artistic style. For totally real images of people i'm still mostly using 1.5 for the flexibility and 'naturalness' of how people look. But Img2Img is really powerful when used right :)

[@RodgerE2472]

I'll give it a shot, thanks!

[@articraftic]

can we do this in comfyui?

[@A_G420]

It almost generates the same face each time?

[@nhuocphong90]

Did you install Nvidia Game Ready Driver or Studio Driver?

[@joywritr]

Are those all open browser tabs at the top? I have 128GB RAM and I wouldn't even attempt that. ;)

[@jason-sk9oi]

Awesome secret sauce!

[@OlivioSarikas]

Hey, someone remembers! I haven't used that line in ages

[@Sarubotai]

It was so simple i missed the trick altogether!

[@Orokana]

I love stable diffusion more than MJ

[@blacksage81]

Yeah, I knew I wasnt trippin, this is the Model swap trick with a twist. Though I will say swapping model during upscale? That's the danger zone baby!

[@timhays2086]

I have been trying to find a way to create photos of my relatives who have died. It is almost impossible to create a LORA or checkpoint to draw them because there are not enough quality images to train on. I have tried Roop but that doesn't seem to work. If you could do a tutorial on this, I would appreciate it greatly. Thank you.

[@garlandpoolfredpool]

Wow!

[@gjohgj]

Also midjourney works nice with generating two face in one photo, vary region one face of that photo and do a side of that same photo. Making a training source for who knows what😂

[@ZiggySA]

The hyperbole is next level. The secondary images do not look significantly better than the base images. Some of them look worse.

[@PeppePascale_]

😂 yeah this is also a good way to empower all old inferences that looks like cr4p, great trick! thanks for your effort in helping the community everyday

[@fastandreiiii]

can you share photos from midjurney plss

[@OlivioSarikas]

I will at them tomorrow as a download link. Not home right now

[@fastandreiiii]

thx for this,you make incredible job with ai

[@fastandreiiii]

@@OlivioSarikas hello where is link for download? thx

[@coloryvr]

...just tried...yes! .....it is GREAT!! Big FANX!

[@blackvx]

Without hands, we can't tell that it is AI 😳

[@magimyster]

meraviglioso!

[@LeonidKotelnikov-jg9fi]

Why the mushroom girl on the opening image doesn't wear sunglasses?

[@yklandares]

its amazing

[@VictorTheLegend]

wish there were a ComfyUI with this workflow, as A1111 no longer runs in my machine :(

[@OlivioSarikas]

you can build this in comfyui with easy