LDSR - best comfyUI upscaler - results at 4k.
Вставка
- Опубліковано 13 лют 2024
- Frustrated by iterative latent upscalers that keep 'messing' with your image? ME TOO! Hence, LDSR - the best for 'professional' use IMHO. / discord
🤗Help keep this channel brutally honest: ko-fi.com/stephantual
Update: 2 weeks in Comfy might as well be 3 years. We now have SUPIR, CCSR and many others to play with. I created videos for those, too, check the playlists!
▬ LINKS REFERENCED ▬▬▬▬▬▬▬▬▬▬▬▬
Workflow: comfyworkflows.com/workflows/...
LDSR: github.com/flowtyone/ComfyUI-...
▬ SOCIALS/CONTACT/HIRE ▬▬▬▬▬▬▬▬▬▬▬▬
Discord: / discord
All Socials: linktr.ee/stephantual
Hire Actual Aliens: www.ursium.ai/ - Наука та технологія
Ok quick update - it's a little more subtle than this - A full upscaler comparison is coming!
Happy to know u are working for them also ;) Thanks
👽
100% 🔥
Can this work with Chainner?
Does this work for Animatediff too?
Were do we import our image?
So is it worth if I have Topaz? Its only upscale or little refiner too?
I'd say if you're a photographer and don't want to bother with nodes and comfy, yes, absolutely. If you are not and you do this for fun, well, it depends on your wallet :) 👽
I don't see the upscale amount? 2x 4x ?? Where is it?
LDSR is a model which is preset at 4x. Run a 2x downscale if you need a smaller image, run it twice for 8x 👽
Definitely on youtube we can't notice how LDSR is better... I love aliens but maybe use other/better examples?
Haha okay that's entirely fair. I'll start generating realistic images for the videos, things you can actually use IRL. Feel free to suggest what type of images or videos you'd like to see generated btw! 👍
@@stephantual Yeah anything realistic, our brains are wired to detect any difference, specially faces. Also maybe compare it with the popular Ultimate SD upscaler too
@@highcollector good points all around! I think you'll be suprised by the results. Maybe I'll be surprised :)
@@highcollector Faces are great and all but i see all upscalers fall apart on architectural or non face scenes. Faces seem to be well represented in all training data, although you do start to notice the same features in many of the generations, especially in user merged checkpoints.
The killer for me here is the minutes of waiting when you can upscale your slop with sdupscale or iterative upscale setups and some lightning model from civitai that matches what you are going for. With turbo and lcm its slightly slower but even better imo. Theres something crappy about lightning for upscaling in my current testing.
@@TR-707 Yeah. I haven't tested Lightning yet for upscaling, but Turbo seems to do great, upscaling fast in just 6-8 steps is a game changer
Those artifacts like little grey squares in browser, I had them too - it's nvidia driver, after latest update it was somehow corrected.
Cheers! Yes it's annoying, I'm very sorry about that. So, I could turn HW accel off in the browser (all chrome based browsers have this), while I record - because I just updated the nvidia drivers to the ones from 3 days ago and there's still there :( It worked for you?
@@stephantual After the update (problem was even mentioned in driver release notes), they appear much less often, but they are still there. First I saw this I thought it was bad VRAM... but happily it's bug 😀 And yes, chrome based only. Zero on FFox. Lucky me, I use FF for Comfy, and Edge for browsing on this W11 computer. IMO disabling HW accel in browser is no option, tried it to hear AIO fans go 'brrrrr' on scrolling.
Thanks for the tip@@Deadgray , I'll use FF to record going forward, hopefully can do everything in 4k from now on :) cheers! 👽
Fooocus has a way better upscale and its less costy on the ram, i have 16gbs ram, 4gb VRAM 3050, give it a try
Check out my SUPIR video, things move VERY fast. It blows everything out of the water ... until something else does :)
That's another good one for upscaling! And Fooocus is free.
oh my god while watching your video i realized im finally getting fat
Don't we all :) 👽Thank you for watching, friend!