Best suggestions for sharp looking face in ComfyUI

[zono50]

I do realize that the model uses a 128x128 resolution, but are there any top level suggestions or expert tips to get the best results using comfyUI? I'm using gfpgan with retinaface_resnet50.

[Gourieff]

ReActor(ComfyUI)+CodeFormer+Upscaler
This girl as a face source: Gourieff/comfyui-reactor-node#19 (comment)

Results:

[Gourieff]

Yep, and I really don't understand why some people complain that they don't have enough resolution of the inswapper model

[FructusVitae]

And what upscaler do you use?

[Gourieff]

NMKD Superscale, Siax

[c4dee]

Hey, i'm getting visible aliasing/edges and smeared details in areas, that are further away from the inpainted faces.

I'm using A1111.

Do you have any suggestions how I can minimize these artifacts? My current solution is to render 2 images. One without reactor and one with the inpainted face. Then I overlay the images in an image edtiting software and carefully delete the areas around the inpainted faces to restore the details.

inpainted face using reactor

[Gourieff]

Hi there!
It's quite strange that you have such a big amount of aliased parts 🤔 what the workflow do you use?
To reduce this effect totally - you can swap via img2img->inpaint with the inpaint mask or automate this process with ComfyUI (by building a workflow with the face mask detection)

[c4dee]

Hey,

my workflow is the following for A1111:

I render an image in text2img, with ReActor enabled. The upscaler in ReActor is set to 4x-UltraSharp.

I also tried "1. Restore Face -> 2. Upscale (-Uncheck- if you want vice versa)" option, but the aliasing remains.

ReActor works much better for me, when I inpaint the faces.

I love the extension and thanks for the feedback.

[Gourieff]

Hi
You can now use it with "Face Mask Correction" option - the result will be very accurate with no pixelation around the face

[2blackbar]

Doods dont suggest some secret hidden fancyass pipeline reverse - this thing should work great from get go and without any changes to anything , pipeline for codeformer is borked and the dev preferes to brush it off and take it personally than fix this and change the obvious dud in the code...
Cmon dev , trash the codeformer pipeline and write it again properly, take a look at roop repo how they keep the quality intact, its not stealing , its learning.
Look at c4dee image above, its so pixelated but its how this repo works ! this is very bad and unacceptable and jaggy af.
Yeah you can brush my comments off and put some shitto smileys on top but this wont fix the issue and blind people with low standards wont see the issue anway so they think me and others are tripping.
Try this one vs roop and take a closer look at the face.
The edge this repo has over roop is speed but the jaggies are currently unacceptable, the default is jaggies, i dont want to tinker around finding SeCrET option to get rid of them.

[Gourieff]

Я показал примеры выше, как работает ReActor, если у тебя не получается - значит, ты просто что-то делаешь не так
Roop в плане замены лица работает по такому же алгоритму, но не умеет ко всему прочему делать правильную маску и имеет кучу прочих изначальных недочётов
Проблема пикселизации актуальная как раз для Roop, в Реакторе это уже давно исправлено
Для свапа мы имеем модель 128х128 пикселей, другой пока нет, поэтому финальное качество зависит от моделей face restoration
Так что утверждение твоё полная чушь. Нравится руп, так и пользуйся им
Будешь не обосновано и дальше гадить в темах, продвигая руп и предъявляя мне что-то абсолютно не по существу да ещё и относительно бесплатного ПО, уйдешь в блок. С Новым годом ;)