Roop vs DFL model comparison

Not really, the stuff we mostly need and which is often missing are all the tricky angles (looking directly up/down) and it's what it struggles with the most.

Yeah, fine for simple videos without blowjobs, we'll see if facials are doable, another test video incoming!

If you mean full video with this girl then this is actualy a full video but only about 10 minutes, it's a commissioned video, solo scene, she does mostly solo, has some hardcore but not much.

Full DFL version here: https://mrdeepfakes.com/video/77887/hailee-steinfeld-22-full-version-dfl

You don't really need to know anything, you just install it, run it, select a photo and video and hit run.

Unless you want better results, if you guys are interested I can do another test and pull out some tricks to improve quality, for that I can indeed make a tutorial.

Now there is: https://mrdeepfakes.com/forums/threads/guide-single-photo-deepfakes-easy-and-quick-literally-one-click.14021/

insight looks better somwhat.

Insightface seems like it does a good job of retaining more features from the face than the original video slut, but obviously has a lot of quality/clarity issues with its output.
I don't like that they pulled the model from GitHub so that they could put it behind yet another shitty Discord bot.

DFL struggles in maintaining likeness when using large datasets (such as mine) with DF-UD model, I think I could do better with AMP model, it would look closer to insightface result. Also the thing is the photo I used for insightface was one of the more recent ones where she is older and has specific makeup and that "look", whereas my DFL dataset has faces that are wider range (since she turned 18 so there are plenty of younger faces and plenty of recent faces in the set) so it has to settle somewhere in between for the look.

You can get drastically different results if you use different picture.

I think you meant it looks more like SRC (source, celebrity) than the DFL one which adapts a lot to the DST (target/porn/the model used), and yeah, it does retain likeness from that one photo but it absolutely ignores facial expressions, it doesn't look where she looks at any given model (DFL does), doesn't close mouth/eyes properly (DFL does), can't handle the dildo in mouth at all (DFL can), masking is iffy (hair is pixelated, DFL is masked perfectly) and clarity is dreadful, but then it's only 128 pixels, with good GPU and DFL you can push for 512 or even 640 which if you have plenty of time, excellent dataset, you could do 4K videos easily, but in reality high res is hard, upscaling is the way to go really so if you guys want to suggest some clip and celebrity, let me know, I'll try to make one that looks better using some post processing and if there is a demand I'll post a little paid (but nothing too expensive) video guide on how to make perfect deepfakes with insightface.