AMD performance for 4.0.0

[Riddle94]

Greetings,

I was looking for optimal settings or some more insight in the software to improve fps render count.
Currently i use around 50% of my RAM, 40% of my GPU, 15% of my processor.
without post processing i get around 8-10fps (5-8 with GFPGAN)
Can certain settings make it that i have more use of the HW i have?

Setup:
WIN 11 64 bit
AMD Radeon 6950XT
i5 13600kF
32 GB RAM

current settings:

[Elise96nl]

look at the usage graph in AMD radeon software. I saw 40 procent aswell on the GPU, but it's just spiking, it hits 40 to 98 in high spike. But indeed only using 5 of my 16gb vram. But im pretty happy with my 4 frames codeformer on 256x

512 and clip2seg are the brutal ones

[C0untFloyd]

Is there an improved quality using the upscaler + codeformer instead of using (native and fast) 128px + codeformer? Really curious, never tried this myself.

[Elise96nl]

with 128px you need high codeformer to get sharp quality results.

with 256px you need less codeformer and you keep more of the orginal faces

but overall withouth any post processing you could have tiny mistakes in the face, or not enough blending. so low codeformer or gpen = preffered for me. but for the tiny things

512px is to high of perf penality and im doing max 1920p video or 2k pics. so most faces arent more then the 256px anyway in this content. when you do a close up headshot in a big picture maybe 512px is worth is.

but atm my results with 256px + codeformer at 0.6 or gpen at 0,4 are amazing, both in video and pictures. pictures are almost always finished products with these settings even in a big batch, without tuning picture for picture with different settings. Do mind for pictures with hair before face, or like half face selfies, phones and hands in front of a part of a face the DFL XSeg is great

for video's gpen is better because of the eyes and mouth. Video's with just a face posing and moving from angle to angle with mouth closed are perfect with no masking and 256px + gpen at 0,4 or 0,5.

if they start talking or head movement and angles are going to be more you gonna need to switch on the clip2seg with tags like "open mouth" so it maskes the whole mouth area from the original input. But if its 2 girls dancing + lipysyncen to a song on tiktok (my usecase haha) the small mouth movement is fixed with 128 or 256px + gpen at 0,60.

so to summerize

pic face close up = 256px or 512 upscale + low codeformer for likeness
pic face away = 256px + codeformer (0,6) or gpen(0,35)
pic half face/phone/hand/hair = 128px or 256 + gpen on 0,4 to 0,6

vid no mouth movement, = 128px + gpen (0,4)
vid mouth movement or closer by = gpen (0,6)
vid but sometimes hand/hair or something in between = DFL XSeg + gpen 0,4
video talking or something else with mouth = clip2seg (slow)

Faceset
But one more thing the faceset input = key to everything
I have facesets with good pictures that are pretty bad, and a few that just work everytime from angles and are just great

the best ones are 2.5/3 mb and contain 8 or 10 pictures. But i cant define what makes some of them so good, and better then the others. Maybe its just some people have easy faces for the software. The one that is the best, containts really bad reso input pics, so its not that.

[C0untFloyd]

Interesting, thank you for your detailed report. I'm mostly avoiding using the enhancers because they change too much of the face identity for my liking. Instead, I often use 512px and number of swapping steps set to 2. There's a certain advantage in using the 512px upscale where I'm sure I'm the only one who knows, because it is linked to the inner technical workings 😉
The swapping process cuts out the original face, resizes it to 128x128 pixel, swaps the face, resizes it to 512x512 pixel, optionally enhances the image and afterwards resizes it to the size of the cut out frame. All the resizing and especially the one down to 128 pixels introduces very noticeable artifacts if you have repeating patterns somewhere around the face (curly hair, some detailed texture). This doesn't happen when using the 512 pixel upscaler, because the downsizing is removed from the process. I know it's a big performance hit but I think it's worth it if you want to fake some tricky GIF with heavy compression artifacts.

Facesets:
In my experience so far, it's important to have similar lighting conditions. Also be consistent, don't mix different age periods, don't mix faces with glasses and without. I think it's a bit like picking images for training a LORA (of course there is no training involved here).

[Elise96nl]

in fact i forgot to give you a big shoutout and thank you. Amazing app. and i hope you and the community will keep this thing going. lot's to improve. AI and stuff like this is going fast. I've you imagine the following video i've made of myself in titanic. is something hollywood could only do with a team of cgi artists, a few years ago. this is something me as a noob did in 30 min on my home computer. It's amazing! Thank you for everything @C0untFloyd

https://streamable.com/1u7n14?src=player-page-share