Realtime Image Generation with SDXL Turbo

[@Zevil]

The sheer speed at which AI is advancing is impressive yet scary. I enjoy watching your videos a lot, Jarod. Thank you for the constant updates

[@Jarods_Journey]

The speed is astounding! Appreciate it:)!

[@Airbender131090]

Yeah. Bit its bad quality and 512x512. When this speed available in 1024x1024 with normal SD XL quality - that would be interesting

[@zengrath]

wow i have to try this it's amazing. got a 4090 myself so i may not have bothered with turbo for lower quality but i never thought that doing this was possible

[@janvollgod7221]

The SDXL is nice for playing around, but the results are not so much what i expect. The progress is great.

[@Kelticfury]

This is a really nifty toy in current state. It is in alpha atm and I am really interested in watching were it goes from here!

[@RaysAiPixelClips]

Was just experimenting with this last night. It's crazy...

[@yoAventus]

This is literally mind blowing LMAO thanks for the content as usual man it's awesome seeing this stuff in action

[@adamstewarton]

This is fenomenal. I hope for a personal Jarvis in the near future that will help us with various tasks

[@Mowgi]

Wow, that is insane!

[@Jarods_Journey]

Too insane :)!

[@doonk6004]

Exactly what the title says! Incredible!

[@RobertJene]

here I was researching what I should do for my next Stable Diffusion video and what do I find, my bro, Jarods Journey

[@swannschilling474]

Its really great for realtime applications and for testing out stuff fast!

[@tukanhamen]

Cant wait for someone to come out with an opensource version of one of those live sketch to AI things. Would be so cool!

[@aymericrichard6931]

It is impressive to test tokens and check which have impact on the prompt

[@dashh3153]

Thanks for the video! Quick question: does your hardware hide the fact that each generation the server reloads the checkpoint or are there configurations that I've adjusted incorrectly?

[@Hooooodad]

Amazing stuff, mate you reckon rtx 4080 16 gb would be a good option on this day and age or should I take the leap and get a 4090

[@Jarods_Journey]

4080 is good, but all depends on your use case. If you need to load larger models, 24gb might be necessary, but you'll need to verify what you wanna do with it. Both SDXL and SDXL turbo work well on the 4080

[@Hooooodad]

@@Jarods_Journey mainly using SDXL , animatediff and new sdxl video model, will look tondo local Lora training too but nothing too crazy

[@lhxperimental]

What GPU did you run it on? Whats your system config?

[@bellynymattos1570]

In description bro

[@Ekopop]

have different seeds for different results at the same time

[@phylippusvonhohenheim3232]

Bro please, you must test this model in your 3060 gpu, this is my gpu too and i really want to know how good it performs in SDXL or is better in google laptop/notbook emulation

[@Jarods_Journey]

Works just fine on a 3060 12gb, generates images in 0.3 seconds

[@fountainbird]

Awesome. I just tried control_after_generate set to increment lol

[@KrishnaKotabhattara]

Amazing content man !

[@RobertJene]

11:24 would like to see a test on a RTX 3060

[@the_nows]

Is there a throttle node available so you can defer the upres loading?

[@ClimbHighWithAI]

is there an option to make pictures from a base image? for like changing faces and creating new characters? or just changing the pose for a specific character?

[@Jarods_Journey]

That would be things like img2img and controlnet, which I'm not sure how turbo works with those workflows.

[@ThatGuyFromDetroit]

@@Jarods_Journey it works pretty great with that kind of stuff, too. i've been using it on Auto1111 (comfyui can run as an extension/tab INSIDE OF auto, which is cool because i mainly only tinker with comfyui for realtime use nowadays), 5-20 steps, cfg 1-2.5, most Samplers seem to work fine on Auto1111 (but not all!) -- seems like (up to a point), giving it more steps/higher resolution, newer samplers, etc, actually does boost the quality! (you can tell when it's too much because the image starts to fall apart, as always) i honestly thought Turbo was gonna be a low-quality gimmick like LCM, but honestly, 1024x1024 (or 768x1152 etc) resolution, 15 steps + 10 steps of hires, and Turbo is almost as good as SDXL at more than twice the speed, AND it listens to my custom-trained LORAs better somehow, too. this is insane. - it might not be 'realtime speed' with my settings cranked up like that, but it's still more than twice as fast as SDXL

[@dhrumil5977]

Hey is this available in fooocus ??

[@alkeryn1700]

i'm curious about using it with the sd upscaler script lol. EDIT: i tried it, does not work well if you use turbo for the upscale as well.

[@Looki2000]

Me right after clicking on this video: "WHAt?! what??! There's no way!!"

[@funkahontas]

This is literally game breaking 🔥🔥🔥🔥🔥🔥🔥🔥

[@JLITZ88]

seems like you are making videos, just need to auto save each change and you got a video. Then the next step would be voice to speech prompt inputs. enless possibilities.

[@watcher_xD]

This is better then mid journey???

[@Airbender131090]

How is this better? This is super fast but quality like mj v3 ( as midjourney was 1 year ago )

[@watcher_xD]

@@Airbender131090 mean midjourny is better?

[@0A01amir]

@@watcher_xD SDXL>Midjourney>DALL-E3>SD1.5>Firefly etc...

[@firasalmeshari4851]

Interesting...

[@archi_designer]

local machine or online?

[@Jarods_Journey]

Local

[@Airbender131090]

PS your 4090 could already generate 20 steps 512x512 image in less than a second 😂 i have 4090 and was not impresed with sd xl turbo

[@Jarods_Journey]

This was not the case when I tested it way back when, but maybe you have some secret sauce I'd need to know of 🤔

[@neey3832]

users with less than 8gb vram: []

[@TPCDAZ]

Quick but poor quality. Any model is quick when you lower the quality and steps to the point of it becoming pretty unusable.

[@KushG420]

But the quick mode is extremely great for taking inspiration where you don't need to wait 10 second for just 1 image

[@ronlawrence99]

Can this be done in SD then linked to oogabooga to give seemless results with a chat bot?

[@Jarods_Journey]

Probably possible, someone would just have to incorporate it