FASTEST Way to Get FULL Quality in 8 Seconds - FLUX TURBO Lora!

[@OlivioSarikas]

#### My Links from the Video #### Get WORKFLOW here: www.patreon.com/posts/flux-turbo-lora-114448319 My Test Results: www.canva.com/design/DAGUN558rXM/IHXsyYzvuEBSrIL11R2izw/edit Flux Turbo Lora: huggingface.co/alimama-creative/FLUX.1-Turbo-Alpha

[@TheGuitarnob]

can i work flux with 11gb vram??¿

[@OlivioSarikas]

@@TheGuitarnob people say they can run it with a 6GB card. but that might be for the smaller gguf models. not sure

[@jovanniagara]

cool workflow olivio !! Very detail and fastt !

[@SkullModder]

hey there, thanks for continuing to make these great videos often!. I use sdxl pony mainly because of it's knowledge of nsfw characters without the use of a lora, do you think pony is still the best for this?.

[@Elwaves2925]

Cheers Olivio, it works well in Forge and for anyone interested - RTX 3060 with 12Gb VRAM RealFlux Dev + lora + Euler and Simple + turbo lora = 1m 5s RealFlux Dev + lora + Flux Realistic (Slow) and Beta + turbo lora = 1m 26s RealFlux Dev + lora + Flux Realistic (Slow) and Beta + no turbo lora = 3m and upwards.

[@OlivioSarikas]

awwesome :)

[@bamit1979]

3060 User here. Thanks for sharing the stats!

[@rickandmortyunofficial8986]

how much RAM are you using?

[@bamit1979]

I'm down to 42 seconds on 3060. I'm using Euler Simple

[@Elwaves2925]

@@bamit1979 You're welcome and that's a great time. I expect I would be down to that time too but in the instances above I had a character lora in use too, so that would have added to the time. I recommend giving the Flux Realistic sampler a go with Beta scheduler (haven't tested others yet), especially if you're using humans. While it doesn't completely get rid of skin shine, it adds in a lot of detail to reduce that plastic look, like pores, vellus hairs and so on. The lizard with snow on it that Olivio did in another vid....no problem.

[@astrovah]

Awesome! Works flawlessly on FORGE, even with GGUF. I'm using a 3090 and i'm getting 15 seconds per image now.

[@Rememberthis-t3j]

Yeah, I really don't like the Comfy spaghetti and prefer to use Forge too.

[@astrovah]

@@thaifalang4064 hahaha I don’t use the 3090 with gguf, it was for testing. I use the fp8 dev.

[@camilovallejo5024]

Nobody really loves comfy but it's just that you have to use it.

[@Gumibeer77]

@@camilovallejo5024I love it 😅

[@duniasainsasrofi]

Thank you Olivio, love from Indonesia 🇮🇩

[@xKingNishi]

thanks for making me keep going down the ai image generation rabbit hole you earned my patreon

[@aegisgfx]

Just FYI, its about 7 minutes per 1024 image on a 1660 with 6g vram, and around 40 seconds on 3060 12g vram

[@rainniu]

Love to see it! Thank you! By the way - have you had trouble with images being a little blurry in Flux? I mean just a tiny bit blurry. It’s enough to be irritating. I’ve played with guidance, CFG, steps, scheduler and more. It’s always just blurry enough to be irritating.

[@OlivioSarikas]

yes, at times they are blury. but a different seed does it

[@hustlecoin]

Love the shirt!

[@helveticafreezes5010]

I'm glad to see you try this out today. I remember yesterday that guy in chat spamming for you to use it.

[@OlivioSarikas]

oh really? i didn't read it, just removed it XD

[@antiplouc]

@@OlivioSarikas yes it was me....thanks for the respect.

[@ChristianIce]

NF4 and Turbo Lora don't work in Comfyui, yet I see it's possible to combine them in Forge.

[@armauploads1034]

Thank you, Olivio, like always! Here is my test result (Forge): Ryzen 7 3700X, 32 GB RAM Nivida RTX 2070 with 8GB VRAM flux1-dev-fp8.safetensors Without Lora, 20 Steps: 88 seconds With Turbo-Lora, 8 Steps: 39 seconds (~56% faster) With Turbo-Lora, 20 Steps: 83 seconds (~5,5% faster)

[@weeliano]

Another banger Olivio! Replicated your Comfy UI workflow quite easily and it delivered excellent results!

[@starbuck1002]

I dont understand. Whats the difference between FLUX Hyper SD 8 step model?

[@demischi3018]

The bnb-nf4 versions already have the clip and vae baked in. no need to load them extra.

[@Gli7chSec]

can you do a video about Animating pictures ( i want to make a music video ) using flux in ForgeUI

[@fooboomoo]

At some point forge removed svd support and now I don't what to use either

[@dsphotos]

hi, i got your workflow from patreon for flux turbo, but only getting noise instead of an image, i have downloaded and put everything where it should be

[@Paulo-ut1li]

I wonder is there's possible to create a super turbo flux workflow with this, it would be awesome.

[@OlivioSarikas]

I will try that tomorrow and update my patreon workflow

[@sorijin]

How does this compare to Hyper Flux?

[@TheAdsys]

I have a hunch, the lora actually didn't work when you were using nf4 models at all. It only worked with gguf but you got slower results. I couldn't get nf4 to work with a lora myself

[@StargateMax]

So it requires a completely different workflow, it won't work on top of a regular ComfyUI workflow?

[@simbarules777]

Thank you 🙂

[@bamit1979]

Woohoo!!! Down to 42 seconds from 2 minutes on 3060!

[@NamNguyen-zj6to]

Nf4 bnb model is already have text encoder...no need load it

[@OlivioSarikas]

at 11 GB? i will try that

[@NamNguyen-zj6to]

Except u want use t5 as 16fp...its will more understand ur prompt

[@OlivioSarikas]

@@NamNguyen-zj6to so, then whats the point XD

[@7nesctv]

Sir, how do I display UI options: SD, XL, FLUX, ALL?

[@ritpop]

Maybe is possible to make a flux lightning.....

[@OlivioSarikas]

That would be awesome :)

[@jovanniagara]

I have never heard of the Alimama company

[@OlivioSarikas]

We never have heared of any open or closed source AI Team, until they published something ;)

[@MilesBellas]

..it's part of the billion dollar Alibaba Group... .

[@kironlau]

@@jovanniagara in Chinese wordings，mama means mother, baba means father though Alibaba is original Arabic words....

[@byujkt3822]

@@kironlau Same with others country mama is mama or mom, baba is bapa or papa

[@RebelliousX]

Thanks

[@greg8909]

i'm no expert but FLUX1 DEV should work with a 4080 and 32gb ram according to the specs. Maybe there is a config issue or an hardware issue if the computer is crashing while running a program.

[@23dsin]

Does ForgeUI support Flux with ControlNet?

[@Elwaves2925]

I think it works with a couple of them (depth and canny maybe?) but the rest are supposed to be releasing this week or next, if I read correctly.

[@23dsin]

@@Elwaves2925 I need Canny ;) thanks!

[@dolboeb-tz4bw]

How to run flux with 8gb 5700xt? 😢

[@nerual_dreming]

Dear friend, did you compare with ByteDance/Hyper-SD Flux Lora on 8 steps? I use it and it works great, I'm surprised you didn't tell about it too.

[@alifrahman9447]

NF4 supports lora!?

[@fooboomoo]

this lora or any lora? nf4 supports loras in general for sure

[@DelusionalInsanity]

This crashes my MacBook M3 Pro with 36Gb memory really quickly indeed. Just 5 seconds, before it becomes irresponsive. On the contrary, original FLUX 1-dev works fine and never crashes, although is slow.

[@rbdesignguy]

I thought you had to run Turbo with predibase for it to work correctly? I guess someone must have cloned and forked it for open source now.. lol

[@knives463]

links to VAE models please?

[@weeliano]

Just google ae.safetensors and the first link is it.

[@jasontaylor4582]

why my comments keep getting deleted?

[@OlivioSarikas]

no idea. youtube removes things automatically. also if you use links or harsh words the spam filter removes them

[@joechip4822]

​@@OlivioSarikas How people simply shrug and accept that YT is punishing simple linking to external resources escapes me. This whole social media thing is a child of the f***ing internet, where linking is built into its DNA right from the start. Trying to cage people into proprietary platforms was late 90ies politics - and we have seen what became of AOL and others with this BS. But obviously the generation afterwards has totally forgotten what the Internet was all about originally...

[@bforartists]

Thanks for the video. What i really hate is the file name chaos. The video talks about flux turbo lora. The file is called FLUX.1-Turbo-Alpha, without any hint towards a lora. And the actual file is called diffusion_pytorch_model.safetensors. Which i need to rename manually now. I really dream about having one explicit name for one file ...

[@mrxeska]

RTX 4090 speeds? My result is about ~6,5s

[@petitpoidferoce]

Why not comparing the turbo model with flux schnell?

[@OlivioSarikas]

Because schnell doesn't give you the full quality.

[@petrimannisto9318]

@@OlivioSarikas What you mean by full quality? I actually prefer Schnell output. Dev output is overly smooth and plastic looking imho.

[@OlivioSarikas]

@@petrimannisto9318 but that's taste. I'm talking about how many errors it makes for example with hands and feet.

[@petrimannisto9318]

@@OlivioSarikas Yep, it's taste. Each to their own, like they say. But I don't think dev is better with hands and feet either.

[@Elwaves2925]

@@petrimannisto9318 Details and prompt understanding are superior with dev compared to schnell. Dev does have that smooth, plastic look but so does schnell, only without the benefits of dev. That look from dev can be decreased with the correct sampler and/or loras. Of course, if you like what schnell outputs, then great stick with it. It did a couple great images for me that dev hasn't been able to reproduce but overall, I find dev so much better.

[@rmar37]

sd 3.5 out

[@poipoi300]

These kinds of LoRA's always have the same issue. The model becomes dumber, loses nuance. Look at how the cat is disproportioned. It's a large cat but where the little girl is holding it, it makes it so the cat is missing half its length. They're not useless though, I may use it for upscaling, we'll see.

[@mik3lang3lo]

Thank you for giving Forge a try, it's what we all use

[@0A01amir]

It's an inferior tech compared to SDXL. the amount of power it needs vs the finished generation is too much. like you need 4x the hardware power and time to be able to generate images that are 10% better and not always 10% better.

[@bxl2012]

Yes, but Olivio has now shown 2 methods to cut down the render time by 1/3. People find tricks and the development will go on. And isn't the tech development of updated PC components the same: firstly ingeniers concentrate on performance, later comes the efficiency?

[@0A01amir]

@@bxl2012 tbh progress kind of stalled. skin textures in flux are terrible, time and resources to generate high quality images is not worth it even with these updates. even nex-gen hardwares aren't capable to run many of these models, given they are going for 8gb 12gb and 16gb for high end cards again. you can call them whatever but wasting more power, wasting more time to be able to do slightly better is no efficiency but the opposite.

[@DrSid42]

There are solid 4 steps models now. 8 steps ? Please.

[@Jpetersson]

No comfy please, forge only!

[@AlbertoCasanova]

This is only my opinion and my taste: I don't like ComfyUI at all and I close your videos when you start talking about ComfyUI and even hide them when I see comfy on titles or thumbnails. Thanks!

[@MSigh]

🤣

[@korreyj]

🤣.but whats the point of sharing this info?

[@AlbertoCasanova]

@@korreyj Well, I like Olivio clips, but dislike ComfyUI a lot. Maybe this is useful for him, if not, no problem.

[@joechip4822]

Same here... with 'ComfyUI only' tutorials Olivio actually should see significant decreases in views in his statistics.

[@ARMASPIRIT]

@@AlbertoCasanova why you dislike comfyui?