Ideas for DFL

[SPT]

Idea #154564879 :

Pause and restart merging.

Sometimes I want to do something other than browse the web or watch some stream. Like playing a game or using Premiere, but I can't because I have 4hrs of merging left which might crash if I do that.

 

[SPT]

Idea #154564880 :

An automatic log file of the last settings used in merging.

I usually copy paste those settings in a .txt file if I had good results, but this time I was in a hurry and forgot to do it on a long video which I separated in 2 parts. I remember most of it but still it would be usefull.

 

[DarkAlchy]

The ability to do multiple trainings by just telling it a new name to use but as it is the .dat for the parameters is hard locked and universal so you have to zip the folder (or move it) then start over which really isn't helpful and almost nullifies having different filenames since no matter what every single one will use the same dims, and resolutions. Maybe make the .dat the same name as the filename?

 

[abubot]

Ideas for XSeg:

Reading the code I was surprised that there is no merging mode "simply use saved (baked) mask".
Mask was calculated from dest landmarks or from XSEG/main model.
It is posible enable direct using dst baked mask?

Ideas for XSeg editor
1. Button for split segment (= create point).
2. Toggle selected polygon full/hollow
   (it may use existing buttons)
3. Create poly from landmarks.
4. Switching frames by click thubmnails. Horisontal scrollbar to quick scroll thumbnails.
5. Switching frames by Left/Right arrows.
6. Copy polys from previous marked frame.
7. Copy/paste polygons thru frames. Move whole poly.
8. Direct convert polygons to baked mask.
9. Import/Export baked mask to/from picture file to fix in external graphical editors.

---

Now the bottleneck of the DFL is face extraction.
Some frames are recognized very poorly even in manual mode and after denoising.

Perhaps a landmark editor can help (need to re-align after that?)?
IMHO, It would also be nice to be able to interpolate landmarks based on the previous and/or next frame).

Are there any plans for landmarks stabilization?

 

[SPT]

High priority thing to work on :
Either face extraction should be more efficient OR manual extraction should be different. In some cases, you try fixing stuff with manual extract and there's nothing you can do, there must be a better way.

 

[androsk]

High priority thing to work on :
Either face extraction should be more efficient OR manual extraction should be different. In some cases, you try fixing stuff with manual extract and there's nothing you can do, there must be a better way.

i did mentioned some time ago about manual extraction, where you can set/put the parts of the face 1 by 1, for the extreme cases when faces cant be detected right, so you can put one eyebrow here, and there, one eye here, nose over there...

 

[1D0F4K35]

Idea #154564879 :

Pause and restart merging.

Sometimes I want to do something other than browse the web or watch some stream. Like playing a game or using Premiere, but I can't because I have 4hrs of merging left which might crash if I do that.

If I'm understanding you right, you want to be able to leave merging half way through, do something else, and then go back to the merging where you left off?
If that's right then I do that regularly by just pressing escape on the converter screen and then 'click any key to continue' on the cmd screen and it saves it at that state.

You just go back to it when your ready, run the convert.bat as if starting from scratch, tell it to use the exisiting setup on the cmd screen, click return on converter screen (once it's opened the image) and it just continues where you left off.

You can even render a .mp4 half way through too in this way and continue it when ever you like.

I'm useless at explaing things but I hope this has all made sense. I work in this way regularly and it works like a dream.
It's obviously been written in a way that enables you to stop and start in this way (and THANK YOU FOR THAT creator)

Please bare in mind that I use DFL 1 but hopefully it's similar with DFL 2 (after all this I'm sorry if thats not the case lol)

 

[SPT]

Idea #154564879 :

Pause and restart merging.

Sometimes I want to do something other than browse the web or watch some stream. Like playing a game or using Premiere, but I can't because I have 4hrs of merging left which might crash if I do that.

If I'm understanding you right, you want to be able to leave merging half way through, do something else, and then go back to the merging where you left off?
If that's right then I do that regularly by just pressing escape on the converter screen and then 'click anywhere to save' on the cmnd screen and it saves it at that state.
You just go back to it when your ready to do so, run the convert.bat as if starting from scratch, tell it to use the exisiting setup on the cmnd screen, click return on converter screen (once it's opened the image) and it just continues where you left off.

You can ever render a .mp4 half way through too in this way and continue it when ever you like.

I'm useless at explaing things but I hope this has all made sense. I work in this way regularly and it works like a dream.
It's obviously been written in a way that enables you to stop and start in this way (and THANK YOU FOR THAT creator)

Gotta bare in mind that I use DFL 1 so maybe you cant do it in DFL 2 (after all that, sorry if thats the case lol)

Thanks for the info. I use DFL 2 but I will try it on my next merging.

 

[1D0F4K35]

Thanks for the info. I use DFL 2 but I will try it on my next merging.

I do appologise, I said it was 'click anywhere to save' on the cmd screen.
It's actually 'click any key to continue'. You knew that anyway though, right?

 

[Grrkin]

Change a setting when starting the trainer, and it goes:

*******************************************************

Loading samples

*******************************************************


I see you have changed your settings since you last ran this model, would you like to save a backup before launching? (y/n)
Y
Saving backup...

 

[vanillaGorilla]

Far down the road, but the holy grail for me would be to live stream a fake.

You'd have to have a model pretrained before going live.

Your computer would have to be able to render the new frame on the fly within a very small window as some lag is always there in live streaming.

One of the benefits to this is often live streaming like on Twitch or whatever, the video itself is small as the content is the king. You wouldn't need a ton of resolution I wouldn't think.

If you had a good deep voice swap and the ability to live deep fake, you could sell that shit to just about anyone making content.

 

[DexterousDubu]

"iperov" I came across this video today:

The tech seems very promising for DFL. Could be used to take otherwise unusable Src faces due to lighting/shadows and make them usuable or more uniform with other photosets or to further train models with and without shadows? Of course I don't really understand what it would take to implement such a feature but just thought I should share it

 

[Fuorissimo]

It would be very useful to be able to save and use your own configuration, recalling it at each training. With the possibility of being able to save several. In this way it is possible to start a training with a configuration, stop it and resume with a second configuration (for example active gan, etc.)

 

[TMBDF]

It would be very useful to be able to save and use your own configuration, recalling it at each training. With the possibility of being able to save several. In this way it is possible to start a training with a configuration, stop it and resume with a second configuration (for example active gan, etc.)

That's not very clever or good idea, you never change more than one parameter so having a bunch of model configuration options for a model would make the process even more complicated, just pressing a key when starting training and changing one parameter at a time is simpler, easier and makes more sense.

 

[Fuorissimo]

You probably did not fully understand the proposal. When training using pretrained set it makes sense to keep the set's settings. Then when the model is defined enough, 3/4 parameters are changed (its in guide , training parameters). It is also much more convenient (and avoids errors) by launching a training by loading options from a file, than pressing 9/10 keys. One is immediate, avoids errors and allows you to have more configurations, the other forces you to do it all over again. I'm sorry you think my idea is stupid. Opinions... :shy:

 

[SPT]

Something I'd like : using XSeg src for width of the face, and XSeg dst for height in merging. Usefull when src has a rounder/larger face than dst. You can approximate this with negative erode, but something like what I describe would be simpler if it was just a merging option.
Or maybe even simpler use dst mask above the eyes and under the mouth, and src mask for the rest.
Or maybe there's a better solution to keep src face shape that escaped me until now because I chose my dst faces better.

 

[Grrkin]

Something I'd like : using XSeg src for width of the face, and XSeg dst for height in merging. Usefull when src has a rounder/larger face than dst. You can approximate this with negative erode, but something like what I describe would be simpler if it was just a merging option.
Or maybe even simpler use dst mask above the eyes and under the mouth, and src mask for the rest.
Or maybe there's a better solution to keep src face shape that escaped me until now because I chose my dst faces better.

This would be freaking amazing.

 

[SPT]

DFDNet integrated in DFL as the new 4.2) data_src util faceset enhance.bat system. Or a possibility to use both DFDN or the old enhancer.

And maybe it could be also integrated in a way that makes it usable on merged pics before conversion into video.

Would this be awesome or super awesome ?

Also some lazy people that I know exist but are absolutely NOT me (I swear ) wouldn't be against having to simply click on a bat file to make it work instead of learning all the process. God damn lazy people ! :s

 

[TheHumbleMrBlue]

I would like to have the option in training to make the mouth a priority so I can get more detail in teeth (it's especially difficult to get detail in the bottom teeth).

 

[honor]

I would like to have the option in training to make the mouth a priority so I can get more detail in teeth (it's especially difficult to get detail in the bottom teeth).

Use this version of DFL - it has mouth Priority and even option for using OLD gan if you want ...