Hyper-SD - Better than SD Turbo & LCM?
Вставка
- Опубліковано 4 жов 2024
- The new Hyper-SD models are FREE and there are THREE ComfyUI workflows to play with! Use the amazing 1-step unet, or speed up existing models by using the LoRAs.
Better than LCM? Take a look for yourself and see!
Want to support the channel?
/ nerdyrodent
Links:
huggingface.co...
== More Stable Diffusion Stuff! ==
Installing Anaconda for MS Windows Beginners - • Anaconda - Python Inst...
Installing ComfyUI for Beginners - • Install Stable Diffusi...
ComfyUI Workflows for Beginners - • ComfyUI Workflow Creat...
Faster Stable Diffusions with the LCM LoRA - • LCM LoRA = Speedy Stab...
Make an Animated, Talking Avatar - • Create your own animat...
Make A Consistent Character in ANY pose - • Reposer = Consistent S... - Наука та технологія
Hyper-SD works well with Krita live generation. 60 rodents per minute. Thank you for sharing!
Nice 👍🏽
That fucking outro sound lol
😉
Los fucking canguros
Thanks for this! Outro song is bomb and you're always great content. I love the ability to quickly generate decent images on one pass.
Suggestion: it would be interesting to see a "low vram" e.g. 6GB workflow taking the 1 step generations with this model and then working those into more detailed images.
Nerdy really is the geezer! Cheers, mate
😉
Works in forge and A1111 alongside controlnet ip adapter face id, tested it in 4 steps SD 1.5
what sampling method?
that outro song is a chune!
ROFL, is it just me or is that outro song worth the price of admission all by itself?! I wouldn't be surprised if people started tuning in regardless of what the future videos will be about, just to check out the new outro. . Keep up the excellent work Nerdy, u crushin it my man...err crushin it, my Furry Nerdy Friend!
😉
The 8 step XL lora works well in A1111, but you need to play with the weight 0.8-1 and Sampling method
Thanks. Yes the 8 Step Sdxl one seems to work pretty okay with Sd Forge also. But nothing else.
@@dreamzdziner8484there is a new lighting model, 3 steps! It works well in forge too CreaPrompt_Lightning_Hyper-SDXL
Can you make a video about HiDiffusion?
Its not out yet for any UI but as soon as it is I also want to see this. Im keeping an eye out and hoping it gets released soon
does it work with automatic 1111?
I wonder how this would perform on a measely 4GB VRAM system like mine
Thank you sir.
Most welcome!
for those w/ a1111, is the setup just download the lora, drop it in the folder, and select it?
or does it also need this comfyui workflow thing?
Hey I can't figure out what the problem is with my animatediff I try to use lcm or hypersd but the video keeps coming out horrible green blue purple colors how do I get normal color videos no one has a video showing animatediff with hypersd lora I'm try animatediff lightning 1step mm with 1 step hyper sd1.5 lora please help me I need a video!.
thank you, sir.
Most welcome!
Great, superfast with IPadapter!
Um.. Is that ozzy singing a British reggae tune at the end? 😂
Could you not do something like, full strength lora on step 1, and then with the lora off on steps 2-4, and get 100% quality?
But does it blend?
Almost certainly!
The closing AI generated song, so very nineties 😅
What's it like vs lightning?
Better 😉
They compared it with 1 step, but even though it didn't follow the prompt well, it still looked the best aesthetically, and usually you would use 4-8 step with lightning and get high quality results fast, plus you kinda showed that 8 step is preferable with hyper-sd anyway. I'll take your word for it though
I tried it for a little bit with A1111, the 8steps can be really useful but it's quite limited on sampling methods -- DDIM works good, others may vary and a few will simply destroy the result. Tried the 1step and it was a mess.
Yup, ddim sgm uniform is the best
Can this be used with Forge as well, or is it only working with ComfyUI?
works on Forge and A1111
Sooo... Better or not better ?
A1111?
the 8 step XL lora works well in A1111
what's the different of lightning and hyper sd?
How to make it run faster ? I want to use with canny CN and webcam image in RT )) I mean I have the same speed 1or 2 fps in sec on 3090
Couldn't get it running in A1111. :-(
Hi, I saw that you choose the Proteus-Rundiffusion model, for SDXL, but they have also a model withoutCLIP...If you run the prompt again, with the same settings, you will get your Rodent Detective! Also (rodent:1.4) works too! 😀 So it's not Proteus, but the version of Proteus. ua-cam.com/video/LAQYZWbmkwA/v-deo.html
Yup - if you’re not getting the results you expect, try switching models!
@@NerdyRodent Thanks for your comment, Nerdy!
Do you compare fp16 is faster ?
Seems interesting, but lightning models already do similar stuff and they are easily compatible with other uis like a1111.
Good job this is so much better than lightning!
Hi. Could you recommend a good video enhancer? Thanks
May I ask the author, where can I download the workflow demonstrated in your video? Thanks
You can get my modified workflows from www.patreon.com/NerdyRodent
1111webUIで使えますか??
Hoi, do you know if comfy and webui forge can do multi model locations? as i tend to model hop a bit, and got such a collection now that i split them to each their drive. One for SD 1.5 and one for SDXL. Is there a easy way/launch command to use multiple drives for model storage?
Symlinks can work wonders for any app 😉
It's only available on comfy? Dear God why?!
maybe because it's so super awesome?
with all respect of course
just bite the bullet and learn it. It's so good.
@@SetMeFree Agreed!
03:56 Trump fish riding a bike!
quality > speed
Additional loras work?
Give ‘em a go!
Yes.
China No.1
I can't express how much i despise pixel art, despite being one of those who used Spectrum-ZX. Who are those weird ppl that fall for this absurdly outdated way of displaying things while having all this technologies?
realism does not equal better. artistic expression does not care nor strive to be necessarily realistic. other wise all art would try to be as realistic as possible
Easy there, It's just modern mosaic art 😊
the udio outrow ẃas epic