Generate Faster With Flux Dev Hyper Loras

[@MonzonMedia]

Man I can't keep up! There is now an 8 step hyper checkpoint available! huggingface.co/ZhenyaYang/flux_1_dev_hyper_8steps_nf4/tree/main

[@MonzonMedia]

I'd love to hear what kind of speeds you are getting with your GPU with these Flux Dev Hyper Lora's.

[@sorijin]

On my RTX 4070 TI Super, 14s for 768x1344 with the 16 step 🫡

[@MonzonMedia]

@@sorijin Nice! Just found out there is and NF4 checkpoint available, only 8step though huggingface.co/ZhenyaYang/flux_1_dev_hyper_8steps_nf4/tree/main

[@AInfectados]

@@sorijin You didn't say: -What GUI you are using. -With or without LORA's

[@SouthbayCreations]

Awesome video, very helpful information! Thank you very much for sharing! 🙌🙌

[@MonzonMedia]

Welcome bud! You’ll be flying with the 8step Lora with your gpu! 😬

[@marhensa]

2:35 The artifacting of the face on Hyper LoRA can often be fixed using Euler Beta, in most cases at least..

[@MonzonMedia]

All these images were using the Euler sampler and beta scheduler. But I hear you, that’s why I use the beta one. For most cases as I said upping the steps helps as well.

[@marhensa]

​@@MonzonMedia I trained a LoRA myself and was going crazy because the result had artifacting. I trained again for another day with different settings and got the same result. I tried again and got the same result. Then I realized it's the effect of Hyper LoRA that causes the artifacting. Yes, Euler Beta seems to minimize this effect, but it's not really that effective. Maybe I'll try increasing the steps as you suggested. Thanks.

[@crazytoolman]

last time i watched your video is on playgroundai channel, wow, you are here

[@MonzonMedia]

Yeah this is my main channel 👍🏼 great to see you here! 🙌🏼

[@freneticfilms7220]

Why this isn't getting merged with some nice other checkpoints?

[@MonzonMedia]

There are a few on civitai but most developers are not doing it because of licensing. That being said there are lots of Lora’s being made. It could also be because its architecture is different than something like SDXL so it’s not as simple or they are just still figuring out how to train it properly. Could be a lot of reasons

[@freneticfilms7220]

@@MonzonMedia which once are merged with this Lora?

[@baheth3elmy16]

Awesome

[@MonzonMedia]

@@baheth3elmy16 appreciate it!

[@AnthonyDev]

Nice, Monzon, do you know if it will be available for schnell gguf q0 too? I have rtx 3060 GB and it's running fine with forge.

[@AnthonyDev]

I mean 6 GB and Q4_0, sorry. 😬

[@MonzonMedia]

I haven't heard anything for Schnell but I did read I think on reddit that Byte Dance were working on a 4 step Flux hyper lora. That being said, things are moving so fast, I wouldn't doubt for maybe someone else to do it.

[@Pusty159]

Flux and similar programs are great tools. But their disadvantage is the lack of a "canvas" function. Just like it was in Playground AI. I'm currently considering choosing something to replace PlaygroundAI. MJ is good but if you look at the price plans it doesn't look the best. I think Leonardo is a more interesting option.

[@MonzonMedia]

Flux is just the model, it can be use on platforms that support it like comfyui, webui forge etc. The canvas function from playground is just a tool that they developed. If you are looking for a local platform and can run it, Invoke AI is the best out there. They also have an online version for a subscription but it's kind of pricey. Leonardo's canvas is probably another option for an online service.

[@NotThatOlivia]

TY!!!

[@MonzonMedia]

You're most welcome my friend! 😊🙌

[@MarcoEscobedo10]

I have rtx force asus 3060 8 gb vram 32 ram. An image 896x1152 + 2 loras = time 1 minute 20 seconds. I have been testing flux_1_dev_hyper_8steps_nf4 , it is faster but image quality is very bad. Now I continue with flux1-dev-bnb-nf4-v2.safetensors in Forge

[@MonzonMedia]

Why are you using 2 lora's? These are meant to be used individually with lower steps? Or are you just saying in general? I find the Hyper NF4 works better with 10-12 steps in some occasions.

[@MarcoEscobedo10]

@@MonzonMedia 1 realism lora + 1 Japanese girls lora + Hyper-FLUX.1-dev-16steps

[@MonzonMedia]

@MarcoEscobedo10 oh I see, that makes sense 👍🏼

[@Ancestor_Numero_Uno]

My ol´ Nvidia 1070 8GB VRAM and 16GB RAM takes 180/sec per iteration. About six times slower than ordinary slow speed. Go figure.

[@MonzonMedia]

Which dev model are you using? 16GB of RAM is probably your bottleneck there.

[@Ancestor_Numero_Uno]

@@MonzonMediaflux1-dev-fp8

[@ronbere]

no gain for me...

[@MonzonMedia]

What’s your system specs? And which dev model are you using?

[@ronbere]

@@MonzonMedia 3080TI 12go Ryzen 9 and 32go ram. 47sec for 20steps in 896*1152

[@MonzonMedia]

These lora's are to be used in 8 and 16 steps respectively. That's the point of them. They are meant to be close to the same quality of the regular dev model but with lower steps.

[@ronbere]

@@MonzonMedia that's right, and I have almost the same rendering time with these loras in 8 or 16 steps.

[@MonzonMedia]

Definitely something not right, with your gpu you should be getting faster speeds. Especially for the 8step lora