DFL Model test DF-D vs DF-UD, Masked Training, Default Masks vs XSeg.
Duration: 3:45
Views: 34 264
Submitted: 1 year ago
Submitted by:
Celebrities:
Maisie Williams
Original Pornstar: Unknown
Description:
Watch Maisie get swapped 5 times, it's like gangbang except it's just another fucking model testing and god did I fucked up here because the results are still inconclusive.
The idea here was to again compare D to UD but then I got side tracked and decided to test variants of masked training being enabled (I'm running this on modded DFL so I can use it with Full Face models, you can only switch it with WF) and also using xseg and default masks to see if there is much difference and I can safely say that??? There isn't? At least I can't tell for certain which one looks better. I also tested style power being on or off with UD but I've made mistake of doing extra training with SP being off on a model that had SP set at 0.1 before so it didn't forget this data and thus result of the no SP are basically identical to the test with SP set to 0.1.
All in all having SP enabled does fix lighting somewhat but it's nothing major.
D seems to be coming out sharper than UD even if I gave UD more training time and higher on average batch size.
No difference in face shape (all were merged with Xseg-prd) between having trained model with XSeg or default masks, also no effect on how face looks.
Masked training does seems to change something about how face look however, hard to describe, if you watch the still frames at the end pay attention to the difference in face shape as well as color matching and light direction, with masked training being disabled face has much more even lighting, in some cases it looks better and matches DST more (1st picture) but in other cases face seems to be lit in places where DST is not (3rd picture).
I think for best results in this case one would use DF-D model, run style power with masked trainign enabled and maybe use XSeg that masks closer to eyebrows... Guess that's another thing to test out.
BTW, if you guys want me to upload these short clips in better quality/better done and without all the text just let me know, otherwise I'll be making something else for an actual release
If you liked this video make sure to hit the like button, add this video to your favorites and subscribe to my profile! (you can then find my videos in your subscription page, upper right corner, next to messages and log out).
I put a lot of work and time into making those edits, please consider supporting me with a donation or video download using tokens (which supports me and the website). To donate to me click on the donate button next to my profile name under any of my videos and select cryptocurrency to find my BTC address or PM me for address.
Also the more given celebrity lookalike gets downloaded with tokens the more content of her I make!
Check my forum thread too where you can vote in the poll and decide what I make next! You can also suggest pornstar matches to celebs I make or plan on making, suggest scenes to edit or just simply request celebs and see what I'm working on right now, talk with me and so on:
//deep.whitecatchel.ru/literotica/forums/thread-request-poll-voting-vote-for-which-deepfake-you-want-to-see
I also make custom fakes of all celebs, if you want to order a public or private video please contact me here on the website or on our forum for details.
The idea here was to again compare D to UD but then I got side tracked and decided to test variants of masked training being enabled (I'm running this on modded DFL so I can use it with Full Face models, you can only switch it with WF) and also using xseg and default masks to see if there is much difference and I can safely say that??? There isn't? At least I can't tell for certain which one looks better. I also tested style power being on or off with UD but I've made mistake of doing extra training with SP being off on a model that had SP set at 0.1 before so it didn't forget this data and thus result of the no SP are basically identical to the test with SP set to 0.1.
All in all having SP enabled does fix lighting somewhat but it's nothing major.
D seems to be coming out sharper than UD even if I gave UD more training time and higher on average batch size.
No difference in face shape (all were merged with Xseg-prd) between having trained model with XSeg or default masks, also no effect on how face looks.
Masked training does seems to change something about how face look however, hard to describe, if you watch the still frames at the end pay attention to the difference in face shape as well as color matching and light direction, with masked training being disabled face has much more even lighting, in some cases it looks better and matches DST more (1st picture) but in other cases face seems to be lit in places where DST is not (3rd picture).
I think for best results in this case one would use DF-D model, run style power with masked trainign enabled and maybe use XSeg that masks closer to eyebrows... Guess that's another thing to test out.
BTW, if you guys want me to upload these short clips in better quality/better done and without all the text just let me know, otherwise I'll be making something else for an actual release
If you liked this video make sure to hit the like button, add this video to your favorites and subscribe to my profile! (you can then find my videos in your subscription page, upper right corner, next to messages and log out).
I put a lot of work and time into making those edits, please consider supporting me with a donation or video download using tokens (which supports me and the website). To donate to me click on the donate button next to my profile name under any of my videos and select cryptocurrency to find my BTC address or PM me for address.
Also the more given celebrity lookalike gets downloaded with tokens the more content of her I make!
Check my forum thread too where you can vote in the poll and decide what I make next! You can also suggest pornstar matches to celebs I make or plan on making, suggest scenes to edit or just simply request celebs and see what I'm working on right now, talk with me and so on:
//deep.whitecatchel.ru/literotica/forums/thread-request-poll-voting-vote-for-which-deepfake-you-want-to-see
I also make custom fakes of all celebs, if you want to order a public or private video please contact me here on the website or on our forum for details.
Download now to support TMBDF
DFL Model test DF-D vs DF-UD, Masked Training, Default Masks vs XSeg.
You currently have 0 tokens available
«Good sirs, who is the model from the original video?»
Reply
«Freya Allan please
One video of her from your side»
Reply
«billie eilish please next!!!!!!»
Reply
Reply to: Sasuke_13
1 year ago«Ok, I see plenty of Billie here, under another video and in few other places so that's what I'll do next. Trying out new pornstar, bigger boobs, she look much better and even face seems to be matching more »
Ответить
Reply to: TMBDF
1 year ago«yeah, billie, we want bigger boobs»
Ответить
«Ellie keeper because there’s not many of her»
Reply
«Make Freya Allan Video please»
Reply
«Thanks for testing the Xseg/No Xseg Mask ! So masked training enabled might give you worse lightning and colors but will fix the YAW mistakes ? Alright.»
Reply
Reply to: Kuzuryushin
1 year ago«Not necessarily, it's kinda odd because so many things change between masked training on/off, the actual mask that is applied (be it default one or xseg one) doesn't actually matter that much, at least not for full face (it does for WF since it covers more area). I'm not sure why there is such big difference between masked training on/off, technically FF always ran (and still does) with no masked training and yet the style powers worked with it, just like they do with WF with masked training being either on or off (as long as correct xseg wf masks are applied) and I only assumed it was there to not have to train rest of the garbage around the face and thus accelerate training (because models prioritizes the masked face area) but apparently there is more to it. Still I think with smaller xseg masks that are closer to those that FANSeg generated the color match on FF should be better, perhaps those huge areas WF XSeg generates can't be used with FF because it might actually not be training as well near the edges of the image as it does in the middle where most of the face usually is. This would explain why most of the face is usually matched fairly well but the foreheads usually are not as well matched (which doesn't matter with WF since whole face get's swapped but it does with FF).»
Ответить
Reply to: TMBDF
1 year ago«I see, anyway thank you for taking on your time to publish these tests.
By the way, have you ever ran the same model and settings twice from 0 to compare the results ? Does it always come up exactly the same ?
I don't think so, using thing like Generative Adversarial Networks, wich is basically letting two IA battle each others should give us a different outcome everytime no ?
Sorry i was just wondering what is the part of randomness in these results. Anyway really good stuff thanks again, it always help to understand better. But this masked training thing leave even more questions haha, i can only test it on WF tho, your DFL is modified. Now i kinda want to run some tests too.»
Ответить
Reply to: TMBDF
1 year ago«Because my understanding of GAN is, it set up a second IA to change the workflow of the main one, the second one is basically trying to bring the most of SRC, that's why it bring sharpness and details but also unwanted stuff if the value is too high or if we let it run for too long. Well that's how i understand it but i'm not quite sure.
And what i don't know is, does things always play themself the same way ?
Like, if we set a limited number of iterations, and we run the same backup twice for, let say 45k with GAN 0.1 or 0.2, does it always end up exactly the same ? Or is there a part of randomness wich lead to differents results in the end ? Maybe you know.»
Ответить
Reply to: Kuzuryushin
1 year ago«No I did not but I'm pretty sure some small differences would always occur, unless you were to train both models at the same time on 2 of the same GPUs and switch all the settings and batch sizes and all always at the same time there will always be some variety which I think is visible in this test I did, certainly non-masked trained DF-UD and DF-D look different but the other 3 look almost the same.»
Ответить
«billie eilish should be next, try some bigger boobs»
Reply