ComfyUI EP05: UpScale AI Image to get High Quality result [Stable Diffusion]
HTML-код
- Опубликовано: 18 авг 2023
- Learn how to upscale AI images using ComfyUI in this easy tutorial.
Dive in and make your images look stunning!
Checkpoint Model AI Angel Mix (V3) : civitai.com/models/104608/ai-...
Positive Prompt:
===============
portrait shot , beautiful young asian idol in coffee shop
(high quality:1.3), (photo realistic:1.2)
Negative Prompt:
===============
text, watermark, NSFW, (worst quality, low quality, normal quality:1.4)
Upscale Model Database : huggingface.co/uwg/upscaler/t...
Suggestion
==========
4x-UltraSharp.pth
4x_NMKD-Superscale-SP_178000_G
Next Episode:
===========
ControlNet
Bro you are like blenderguru of Comfy UI! Such a great video ❤
I've been playing with Reactor and now trying to add detail to the smooth airbrushed faces. This series is really helping me, though the face restore models are driving me nuts. One adds skin detail but wonky eyes, the other has good eyes but less skin pores. Anyway, on with the series, see if you introduce better ways :).
Your tutorials have really explained some basic core concepts of how AI works that too many others skim through without explanation. Really appreciate that.
Very informative, thanks :D
ติดตามผลงานครับพี่ ขอบคุณสำหรับข้อมูลดีดีครับ
Thank 😊 you for the 📹 🎥 📻 📸!
Bro is MVP! Much luv
helpful video ❤
Waiting for this, Thanks
Thank You
I only recently started using AI text to image ( i am slow i know ;) ) but thank you for your tutorials they are a great help for me to learn and you teach it so well.
Thanks for the tutorial.
I did some testing on my low end GPU with UU:
I turned a 450x450 image to a 1800x1800 (4x), it created 16tiles and took about 400 seconds to complete.
Then I used a random Upscaling program called Hitpaw (paid program):
I turned the same 450x450 image to a 1800x1800 (4x), it took 3 seconds.
Result: Final quality was basicaly the same while Hitpaw was a bit sharper and kept more details than UltimateUpscaling.
quality of UltimateSDUpscale should be better... so strange
@@AIAngelGallery Well that's Paid vs Free. paid stuff most of the times get updates, hire talents and are optimized etc... for example FaceSwap, we are stuck with inaswapper 128x128 low quality while the dude that made it is now working for MidJourney and released it's 512x512 on that platform, that mean 4 times the detail on a simple face swap nod/model.
All we can do is to hope for more AI leaks and software codes to be avaliable to pulic to be able to do the same stuff while fighting with a bunch of errors, unoptimized crap and sometimes unsafe apps.
Can you do a video to make an image clearer/sharper more detailed.
Sometimes I get an image and I like the general look and poses but I am not able to get a high quality version of it without major weird changes.
ขอบคุณมากครับ🙏🏼🙏🏼🙏🏼
ทำ lora กับ controlnet ไม่ทัน แค่อธิบายหลักการอัปเสกลก็ 20 นาทีแล้วครับ 55
เป็นกำลังใจให้ครับ
ขอบคุณมากครับ
ค่อยปรับปรุงเรื่องการออกเสียง สำเนียง ให้เหมือนเจ้าของภาษานะครับ ไม่ยากทำได้แน่นอนครับ
@@pokpak9966 นั่นแหละครับ สาเหตุที่ผมต้องมาทำคลิปภาษาอังกฤษครับ ไม่งั้นผมไม่มีโอกาสได้พูดเลย
Why do my images turn out opaque? :(
This was very helpful, but sadly could not convince me :)
In A1111, I can have Face restore and then easily do another pass with the same Model to upscale (I do 1.75x) and adjust if I still want Face Restore.
btw: the Sliders are so much easier to work with than this tins Number Bars and different Modules.
Well, then I give the 1080p Picture a 2x Upscale and end up with a 4k Image.
I think Comfy would work good but it is too spread out over so many different/unnecessary Modules, and it lacks features like I spoke about.
It looks to be faster, but it only wins marginally and with a loss in Picture Quality. There have been some very good Pictures created by Comfy, but I do not know if they have been just luck generations or if Comfy can really get out better results overall. But yea... after all, the Upscaled Image looks way worse than through A1111.
Maybe I give it another shot tomorrow and exactly follow your build and only use SDXL. SDXL seems not so much affected by bad Face Generation.
Thanks for the Guide!
Really was helpful to see that seemingly I didn't do much wrong I will see tomorrow exactly :)
The best method for fixing face is using Face Detailers nodes which i will cover in future episodes
@@AIAngelGallery yea... tried that and all the other nonsense I could get my hands on.
Nothing works in "Comfy"UI.
How this can be called Comfy is beyond me! Import other peoples Workflows sounds great (AND SHOULD HAVE WORKED) until you tried out 10 and not one of them is even classifiable as "good". Most of them are broken in the first place and need fixing until you come to the same result = this is waaaaaaaaaaaay worse than A1111.
Even the simplest Setup with Upscaling results in Mosaic Pictures even tho I did everything exactly like you did xD
Just broken!
@@fr0zen1isshadowbanned99 if that so, you can use other ui such as automatic1111.
.
For me Comfy can provide some automation and customization that other tool cannot do.
@@AIAngelGallery That's why I tried it :) and the better Generation times too ^^
Just that it refuses to do any good work at all xD I guess one can do objects and Landscapes etc. with it, but that's not why I'm after it ^^
@@fr0zen1isshadowbanned99 it seems to be that you have not spent enough time with ComfyUI - My experience is the opposite. Comfyui is more flexible and give more options to play and create. A111 is way too pre-defined, and sloooow. I use both - however I find myself using more and more ComfyUI.
จะสอน comfy เพิ่มอีกไหมครับ
ให้ทำได้หลายๆ แบบ
สอนไปเรื่อยๆ ครับ มีอีกเยอะ
รออยู่ครับ ผมมือใหม่พึ่งเล่นได้ 2 วันครับ lora ยังไม่เป้นเลยครับ
@@AIAngelGallery
ลงเมเนเจอร์แล้ว โปรแกรมไม่รัน อะครับ ต้องแก้ยังงัยครับ
ลอง git pull update comfy ด้วย manager ด้วยครับ
ขอบคุณครับ สุดยอดครับ Thanks for your contribution.
ยินดีครับ