i'm going to be a moviestar!

[Elise96nl]

So i've put my face in a big movie (small role), looks really good. manuel clip2seg, 3 runs, frames extracted etc, after i did some noise reduction, deflicker, and other davinci resolve tools. :)

in the meantime, i'm a pretty big uploader on a big platform. So i hijacked the original movie, and made the version with my face in it the official release. ATM 20k downloads, so now i have a little regret, but fun to know a lot of people are going to see me as a moviestar. :p Thanks to @C0untFloyd and roop unleashed!

[RoopUser]

Congrats. You did it.

[C0untFloyd]

Oookay just don't forget that this is all fake 😉

[Elise96nl]

Just watched the movie together with family. Hilarious, it is my best project to date. The result is amazing and terrifying at the same time.

I get the reluctance from the inswapper guys for releasing the 256 or 512 models to the public. You can see its upscaled from 128 on a 1080+ hollywood movie because of the pores and fine lines in the face that are missing in the low res first swap.

With a 256 or 512 model this is going to be something that would be dangerous when done right.

I red an articel about the movie The Irishman. 30 million for CGI and making robert deniro and al pacino look young again. It was a 500 person team. I just redit a scene of that movie in 4K. with a 10 picture face set. I dont have the fine details skill work to fix the last glitches and color grading. But even after 10 min my scene looks beter then the movie.

you @C0untFloyd did an amazing job keeping this tool updated and fresh. But sadly we are bound by the inswapper ofcourse. so next step would be a higher core swap model? or a detail ai generator after swap. To mimick fine details like pores and freckles, and lines. But i quess the second solution would be just as much work as greating your own inswapper like model. I heard inswapper sells their 512 model to goverments and profesional studio's etc for 100K+

some 500 VFX technicians in postproduction facilities in Vancouver, Ireland, China, and India, as well as in Industrial Light & Magic’s San Francisco headquarters, labored for eight months to finish the footage, giving each sequence in which the actors appear de-aged a thorough once-over to remove as much uncanny-valley artificiality as possible.

[Elise96nl]

one more question. how does or upscale work atm. Because without a post processor i can see little difference between the 128 or the 512 option. Im using DML on AMD. I think there might be some issue with the upscaler model used on DML. because i think there isnt a lot of difference.

[C0untFloyd]

Glad you enjoy you it. There are higher resolution faceswap models, you can try them out in face fusion. However IMO none of them comes close to the quality of the one from insightface.
I know I can't compete with the big guys and I don't have the time nor resources to train a model for a better resolution. For me the next steps are:

• Complete Head swap. There are lots of face swappers now but none allows zero-shot head swapping. This will not necessarily be included into roop but another repo. I'm currently tinkering with this https://github.com/C0untFloyd/HairTransfer
• Transfer of eye glasses, freckles etc. This is closely related to the one before because it is all about StyleGans -> https://github.com/williamyang1991/StyleGANEX
• Face editing, expression editing. I'm already contributing to https://github.com/KwaiVGI/LivePortrait

DML: you picked the only upscaler which still uses pytorch 😄and I'm unsure if that works 100% on non-cuda systems. Do the other upscalers work? Did you try GPEN for example? Or are you asking about the pixel upscale? That one slices and dices the input face and presents different slices it to the 128x128 inswapper. The results are then combined into the final result, sort of subsampling it.

[Elise96nl]

you are confused with DMD i quess. Because DML is the execution provider used on AMD, DMD is indeed the only upscaler that doesnt work on my system. GPEN is great. I was talking about the pixel upscale. i tried facefusion, your app is beter. And there also the 128 inswapper is still the best, and pixel upscale changes only a little bit, ofcourse you cant make wonders from 128 to 2048. :P

i looked into simswap etc, but the problem is the dataset. 3 million pics+ sorted, and cropped, and tagged. thats a lot of work.
for the simswap 512 etc they used vgg_face2 hq. and that is just the original 128px vgg-face2 run trough gfpgan to upscale. So thats why its worse then the 128 inswapper model.

we need some hacker to get us the advanced face-swapping models: inswapper_cyn and inswapper_dax :P maybe on the darkweb haha. we shall see.

thanks for the update!