LCM LoRA = Speedy Stable Diffusion!
HTML-код
- Опубликовано: 13 май 2024
- Yes, you read it right - a LoRA which actually helps you generate images FASTER! In just 1 SECOND (4 steps) you can generate a 1024x1024 image, PLUS - being a LoRA - it works with any model, IMG2IMG, Inpainting, ControlNet, etc, etc!
Works with stable diffusion 1.5, SDXL and SSD 1B.
Update! For A1111 without lcm, you can lower the lora strength to 0.5 for a better generation on SDXL!
Enjoy :)
Huggingface Blog Post:
huggingface.co/blog/lcm_lora
LCM Collection:
huggingface.co/collections/la...
SSD-1B - Up to 60% faster than base SDXL and uses less VRAM!
• Generate up to 60% fas...
Workflows:
github.com/nerdyrodent/AVeryC...
Stable Diffusion - Face + Pose + Clothing - NO training required!
• Stable Diffusion - Fac...
== More Stable Diffusion Stuff! ==
* Installing Python for MS Windows Beginners - • Anaconda - Python Inst...
* Make an ANIMATED Stable Diffusion generated avatar! - • Create your own animat...
* Dreambooth Playlist - • Stable Diffusion Dream...
== Chapters ==
0:00 LCM LoRA Introduction
2:45 LCM LoRA for any SD1.5 checkpoint!
5:00 LCM LoRA for any SD1.5 checkpoint (A1111)
6:42 LCM LoRA for any SDXL checkpoint!
8:58 LCM LoRA for any SSD-1B checkpoint!
10:36 LCM LoRA with AnimateDiff! 
Наука
As a good student, we sit on the front row, learning from Nerdy Rodent! Woot woot! 🤭 Looking good.
That's truly a significant advancement, and it will likely be refined in the future to produce even better results. Imagine these Lora/AIs becoming faster. We could develop AI filters that transform entire games into photorealism, including old classics. This might happen sooner than we think.
Yup. They’ve only been out a couple of days now, so I expect loads of stuff in the next 48 hours 😆
@@NerdyRodent Have you already tested the new Nvidia plugin TensorRT for Stable Diffusion? It is supposed to double the speed.
Loving your deep dives into this world man!
I was just looking for a way to use it in A1111, and you are the first one to talk about it. Thank you!
already having fun with those loras, anyway I am in love with your tutorials. They are so entertaining, love your work!
Glad you like them!
Truly amazing 👏
Between this and SSD-1B, Stable Diffusion is hitting it out of the park right now..
Many thanks for this tip.
I wanted to try out a short animation yesterday with Controlnet Tiles, LineArt and TemporalNet (auto1111). A single render image took 25 minutes. With this LORA it was only 5 minutes per render. So yes, it reduces the render time for video enormously. There is not much difference when generating images. At least on my Mac M1 8GB RAM
Nice!
Very cool! Thanks for sharing
This changes everything. If there is a generation site that allows this lora, your credits will go quite the distance :D thanks for the video.
Good video, LCM definitely works well at 3 steps, cfg 1.5 and dpm sde++ karras. However in auto1111 even, without the lora if you set the cfg really low, like 1,2,3 ... lots of samplers work at four steps. For four steps, you can try DPM2 at cfg 1.5, DPM2 a at cfg 1.5, DPM++ 3M SDE at cfg 1.5, DPM2 Karras at cfg 2, DPM++ 2S a Karras at cfg 2, DPM++ SDE Karras at cfg 3. They obviously get better with more steps, but they're usable at 4 or 5 steps.
Those are 2-step samplers. They do 2 steps for every called for steps. That's why they're slower.
I’m sure A1111 will get proper LCM support soon!
Fantastic video my friend !
😀
Mind blown.
Yo. Awesome Video! do you have a link to the comparison graph you made? cant find it on your github
WHAT???? can't wait to try!
skip it, you'd only be wasting your time. the quality is trash at best.
Ur voice is crazy smooth
What prompt did u use for it?
Rodent gentlemen with a curly moustache 😉
Is the LCM animate diff workflow the same? I cant seem to find that one on your Github :)
Have you tried FreeU? with that an UniPC at even 3 steps I'd say I'm getting equivalent or even better results even without the lcm-lora. of course i did fiddle with the FreeU settings
Nice!
do you need 12GB vram to use IP adapter? i can't seem to use it on 8GB
Nice
that bike for sale anywhere near massachusetts mayhaps?
Even regular sdxl seems to only work well for me between 8-12 steps and low control
I'm so new to this. I don't understand how to use it in automatic 1111. Do i just use it as a lora? What weight do i set it to?
Has anyone else put the LCM SDXL Lora into the Lora folder in automatic1111, but then it still doesn't show up in the SDXL Lora list to select with the rest of the Loras?
Yep, only visible in 1.5 models, strangely
Hi guys, It seems like I can't just run it on mac, do you know what changes I need to do or any resources? I couldn't find any on hf or github
Great video, last week sdxl was taking me 2mn+ to generate one image, now 17sec....
Amazing!
Is it normal that I can't see the Lora in my list of loras in A1111? I downlaoded the sdxl lcm lora and I'm using it with an sdxl model. Many other loras show up but not the LCM one.. even after I refreshed and restarted A1111 many times.
Ow boy, neat
Ikr!
I'm attempting to use the workflow you provided, but I cannot find the ModelSamplingDiscrete node, where is it from?
You’ve got an old version of ComfyUI
@@NerdyRodent Thank you so much!
@@NerdyRodent I have the same issue. Updated Comfy via the Manager, but still can't find the ModelSamplingDiscrete node... is there a different way to update Comfy without completely reinstalling?
nevermind... there was a bug that would not let me update and I had to do a complete reinstall and see all the new stuff now
The future is here!
Is there a source for the grid you are showing in the video?
It’s just an x/y grid of steps Vs samplers from A1111
Actually this might be pretty good when is comes to video creation which usually takes a lot of time.
Also does this Lora works with upscaling?
Seems to work with everything so far!
@@NerdyRodent ok that's insane.
Just been playing with this, my potato 4GB card is magically transformed! Which is bad for my health, as I won't be able to go for long walks while it renders, any more. 😁
I'm not sure I'm downloading the right files. Is it the pytorch_lora_weights.safetensors file? Why are all three named the same?
Yes, as mentioned in the video, they all have the same name
@@NerdyRodent Sorry I missed that. Thank you :)
What about riffusion?
Interesting idea - give it a go and let me know!
animatediff LCM next with the new SDXL version
Ooo, nice!
Which sdxl new version? Is it the one mentioned at the beginning of the video?
Anyone else gets very blurry images using this? I mean, its fast, thats nice, but I think about 90% of the images I made using the lora was very blurry. Does this only happen on my side?
I'm using SD 1.5 on A1111 btw.
Same problem here. I suppose fresh AUTOMATIC1111 install might solve the problem... or not.
For Macs only works with cpu so far, so no good to use this lcm loras :(
58s for 4 steps in a Macbook pro 14" 16Gb Ram 2020
Is that 10x faster like they say?
What gpu do you use?
It’s a 3090 so if you’ve got a 40 series card expect it to fly!
👋
👋
I like your tutorials. I can't stand the Comfy elitism that's all around all of a sudden....but at least you managed to -almost- avoid take a dig at those who prefer simplicity.....
I switched to start with as SDXL just failed on A1111, now I’m addicted 😆
@@NerdyRodentYea, totally understandable. Developer was hired by SAI and it definitely caters to the coder/developer types. I have no prob w/ it, and honestly was never all in on A1111, but I'm betting "Automatic" himself is probably the type of person who'd use Comfy and piss on A1111 & those who still use it........maybe that's why there haven't been any pushed/public updates in over a month now :-/ The Comfy dude's app tag line is a bit cocky too... :-p
Anyhow, back to trying to figure out AnimateDff ha :)
@@carlingo3191 heh. If you get any decent AnimateDiff stuff let me know as I’m playing around with that too!
lol I tried SSD-1B-A1111 model in auto1111 (obviously), it took almost 4min to load and the result was slow (around 25sec/it) and very ugly haha bad idea, don't do it guys xD No issue in ComfyUI
Is it possible to use LCM with ip-adapter
Just like I show in the video you mean?
Love your videos! however my pictures always come out grainy and without any detailed background after following your example
First 🐭🐭
W00t!
doesn't make a difference with my 3070 and automatic1111
My automatic1111 doesn't even see it in the browser.
@@generalawareness101 did you press "refresh" in lora menu ?
Surely did.@@666Counterforce
*cries in 1050 ti*
You get 9 iterations per 1 second, I get 1 iteration per 9 seconds lol.
So down to around 36 seconds, nice!
Tried it... The loss in quality is not worth the time you save
I see no difference. 14 second - 5 steps - 1024x1024 with LCM and without LCM. GPU: 3060 12Vram
Generating TRASH quality 10x faster lol
This is crazy isnt it? Rest of my PC is new but my olde 980ti is still kicking! still does all games I play 60fps 1080p & now gens SD in mere seconds, I cant even use xformers because even though I compiled a wheel it was slower than without, as others also reported, so... i've been crawling with like 20 to 40 second gens for 512x768 images, now its single digit seconds, 3 to 7 depending on various factors. Crazy!
Zooom! 😀