@deeppppp @kolessios
Hey @deeppppp , out of curiosity why do you think the new checkpoints aren't giving the desired results yet? Does the variety of new angles in the training set just require more time to train? Is there some other approach you're taking besides more training time to improve the results?
it's a learning process, where every step takes weeks before you can see the results.
I'm introducing new code etc. because my goal is to be able to improve and go beyond the deepnude results.
I've had access to a Titan RTX now for 3 months and in that time made several revisions.. I'm very confident I'll get there, it just takes time and with the corona virus hitting hard.. I can't say when it will be ready, just rest assured work is done on it and results will become available :)
I'm introducing new code etc. because my goal is to be able to improve and go beyond the deepnude results.
I've had access to a Titan RTX now for 3 months and in that time made several revisions.. I'm very confident I'll get there, it just takes time and with the corona virus hitting hard.. I can't say when it will be ready, just rest assured work is done on it and results will become available :)
Maybe I'm too stupid to RTFM but I coulnd't find it so here we go: How can I nudify manually edited Maskfins? So if I understand correctly, I should be able to export generatd Masfins to photoshop, edit them accordingly and then feed it back to DreamTime. However I don't understand how the last part works. Any help would be kindly appreciated :)
@ubolt11_gitlab if you want to use dreamtime, in preferences you can choose “nudify and maskfin” and “nudify with maskfin”, so if you have already a maskfin, you must choose the last option. Then if you want to do all the process manually, you can execute dreampower via powershell, there’s a tutorial somewhere in the main website. If you cant’ find the tutorial, tell me, I have a document saved in my pc
Ok guys, I made a written tutorial on pastebin, this is the link: https://pastebin.com/WJPvLGAa if you don't understand something, I'm sorry for my bad english, but don't be afraid to ask me explanations
I actually wrote a tutorial and finished about the same time as @Piergiox . I'll share mine also as we probably do things a bit differently.
https://pastebin.com/swJNnfeJ
https://pastebin.com/swJNnfeJ
I recommend you do all your mask editing on MaskdetToMaskfin, not MaskrefToMaskdet. Maskfin is much more precise, and in my experience the alg will sometimes make some major changes when generating a MaskdetToMaskfin from the MaskrefToMaskdet, which if you want precision then you probably don't want many changes happening.
A little faster method is this 1. Create a New folder on Desktop and rename it to 1 (for example) 2. In the folder put your photo 3. Open Command Prompt 4. Paste cd C:\Users\ (Your Username)\AppData\Roaming\dreamtime\dreampower 4. If your photo is 512x512 paste dreampower.exe run -i C:\Users\ (Your Username)\Desktop\1\1.jpg -a C:\Users\ (Your Usename)\Desktop\1 -s 0:5 5. If your photo is not 512x512 paste dreampower.exe run -i C:\Users\ (your Username\Desktop\1\1.jpg -a C:\Users\ (Your Username\Desktop\1 --auto-resize -s 0:5
If you want to edit the masks first run steps 0:4, edit the MaskdetToMaskfin photo and the run step 5:5
Also in Command Prompt cd stands for change directory
For your convenience don't use complicated folder names and rename your photos to numerical style like 1,2,3 etc
@Piergiox I discovered that you can do whatever you want to step 4 and it doesn't matter what the previous steps have in them. The masks generated step 0 through 3 could be totally off, but if you input a good mask manually in step 4 then the output will be fine
@xrebel1 If you keep all of your inputs in the same folder (Samples in my tutorial) then you don't have to edit the folder path in the command, just the image name. Same goes for output. I only have to change the image name and steps when moving from one image to another
Or where to look for that info?
@deeppppp This might be a bit outside of what you are currently doing but would it be possible to make a programme that will completely create the body from scratch with any photos of the person or multiple being used to get the face correct. This could somewhat solve the issue of not having photos that are ever going to work with the programme. I'm not saying this because I'm expecting it to be done at any point in the near future, just wondering if its a possibility.
@Rythoka Read through the information here to get a basic understanding of what the mask does.
https://github.com/dreamnettech/dreampower
https://github.com/dreamnettech/dreampower
@Lew30323 There is a basic principle about the algorithm that DreamTime/DreamPower uses. The Generative Adversarial Networks (GANs), a machine learning technology is based on a set of data (in Dreamtime/DreamPower these data are photos) which is used to train the GAN to produce realistic body assets. GANs can't produce results from nothing so a training set of data must be used. Your comment is a bit vague of what you want to achieve but the main goal here is to train the GAN (the algorithm) to produce realistic body assets based on a set of data (photos). I suggest you read a bit about GANs and how they work so you can have a better understanding of the machine learning techology.
@xrebel1 I think I may have explained it badly, what i mean is that instead of relying on the masking to detect what needs to be replaced it would create a basic outline of the entire body and render the entire thing, instead of just the places that need to be replaced, yes it would be alot more work but it would allow you to work with almost any photo as an imput. Basically, instead of only masking the sections that it is currently doing it could mask each and every part seperately, at the moment it is impossible to work on fully clothed photos because it doesnt have enough data to go off, so why not train it to just create the entire body off the skin tone that is visible in the face. On another note, there is currently technology to create entire bodys automatically, there is also technology to place anyones face onto an existing person, why not put the two together? Throw in a load of photos of a person and generate whatever you want off that.
I know im saying that like it is easy, I know its not, like i said before, im just wondering if it is possible, im not expecting it to happen any time soon.
hmm... bouncing off the recent topics, would there be any interesting/significant way to leverage something like this: https://topazlabs.com/mask-ai/