- Jun 22, 2019
- 30
- 4
- Thread starter forbidden101v
- Start date
- Aug 25, 2017
- 2,490
- 3,748
Links, plural? The game they're talking about only has one link.
- Jul 17, 2017
- 1,981
- 1,716
Dear fellow men of culture.
Been playing with stable diffusion for a month or so, I at least think I know what I'm doing half the time.
I'm several hours down the rabbit hole of img2img. I've been practicing rotoscoping, which you can think of as 'video2video', where the output is just filtered by the AI with whatever bias you give it in the prompt. This is time consuming, and hard to get right as it is. Decensoring 1,400 frame by frame images is a whole other ball game.
This is what I've accomplished so far.
Side by side for comparison, results are not yet that appreciable.
I'm yet to find a model, denoising strength, and proper everything else to get this good enough.
Biggest issue right now, is that the dick tends to change completely from one frame to the next, there's no real consistency between each image. And if the denoiser is set too high, the whole image begins to change from the original. If the denoiser is set too low, nothing happens, ie, no decensor.
The only other option is to manually inpaint all 1,400 images by hand, which would probably take me weeks or months.
Currently there is no method to "batch inpaint" as far as I can see.
Any chads out there want to take on the challenge, please do.
I'm probably going to give up pretty soon, what I want to achieve is too niche for the tech I'm using.
Been playing with stable diffusion for a month or so, I at least think I know what I'm doing half the time.
I'm several hours down the rabbit hole of img2img. I've been practicing rotoscoping, which you can think of as 'video2video', where the output is just filtered by the AI with whatever bias you give it in the prompt. This is time consuming, and hard to get right as it is. Decensoring 1,400 frame by frame images is a whole other ball game.
This is what I've accomplished so far.
Side by side for comparison, results are not yet that appreciable.
I'm yet to find a model, denoising strength, and proper everything else to get this good enough.
Biggest issue right now, is that the dick tends to change completely from one frame to the next, there's no real consistency between each image. And if the denoiser is set too high, the whole image begins to change from the original. If the denoiser is set too low, nothing happens, ie, no decensor.
The only other option is to manually inpaint all 1,400 images by hand, which would probably take me weeks or months.
Currently there is no method to "batch inpaint" as far as I can see.
Any chads out there want to take on the challenge, please do.
I'm probably going to give up pretty soon, what I want to achieve is too niche for the tech I'm using.
- Dec 13, 2017
- 345
- 328
You don't have permission to view the spoiler content.
Log in or register now.
instead of trying decensoring it?
- Jul 17, 2017
- 1,981
- 1,716
Have you played this game? It's not 3D.
Regardless, that sounds like even more work, because you would need to crop around the hands, tongue, and mouth for all 1,400 images. Nah, automation is the way to go here, manually editing all of the images is insanity.
- Aug 25, 2017
- 2,490
- 3,748
You can't like automate it to crop the dick out, and then put the dick back in separately, with everything else layered on top?
JK is literally pronounced as "joshi kousei (unless I can't spell latinised Japanese)" and is translated as "girl highschooler" or just "highschooler", basically a fresh-faced hooker.
- Jul 17, 2017
- 1,981
- 1,716
That's inpainting in a nutshell, but two issues there. One is that each image requires you to draw a new mask. That's what the AI reconstructs based on the prompt versus redrawing the entire image. Hence, you can't automate the process, there's no "batch inpaint".
So that's 1,400 masks to draw over each frame of dick, and only 1,400 if the Ai actually makes what I want on the first try every time. Same problem as above too, each dick will look a bit different in each frame.
I've already tried this as well, if you only mask the dick, the AI can't tell where the dick is actually supposed to be. If you mask more than the dick, most of the original surrounding details are lost in reconstruction.
There is a method called full resolution inpainting, which inpaints the masked area, but just pastes the newly created image over the top. This leaves a ton of artifacts on the edges though in my experience.
I'm pretty much at a loss in moving forward. I'd love to demosaic this game, it's been on my shit list for years.
Google just released a new AI video generation tool called "Imagen video" that could solve your issue in the future since it is much better at remembering the context of the earlier frames so the object (in this case the dick) doesn't change completely from one frame to the next. I think soon we could have a tool that could properly decensor frame by frame in the way you are trying to do.
- Aug 25, 2017
- 2,490
- 3,748
- Jul 17, 2017
- 1,981
- 1,716
Read somewhere that some successful attempts at GIF editing are accomplished by putting all of the frames into a single image sheet, and forcing SD to process the entire sheet. As opposed to running each frame separately, this would provide a consistent result for one animation.
I've made some attempts again using this method. The "bl_bl" folder has 30 images, which I've made a sheet that is 3,840x5680. Trying to process it using upscaling tiles, but upscaling has been giving me nothing but issues, and my webui keeps crashing. I may as well stop here. Because there are several folders with 175-190 frames each.
So... merging one sheet with 180 images will not only take forever, but the resulting image sheet would be around 7,680x17,040. I don't have the GPU power to push through an image that size. Not to mention the amount of manual work involved with creating the whole sheet, processing, then dissembling the sheet back into individual frames.
With exactly 1,424 images to be uncensored, I think we still need a better process to come along.
I've made some attempts again using this method. The "bl_bl" folder has 30 images, which I've made a sheet that is 3,840x5680. Trying to process it using upscaling tiles, but upscaling has been giving me nothing but issues, and my webui keeps crashing. I may as well stop here. Because there are several folders with 175-190 frames each.
So... merging one sheet with 180 images will not only take forever, but the resulting image sheet would be around 7,680x17,040. I don't have the GPU power to push through an image that size. Not to mention the amount of manual work involved with creating the whole sheet, processing, then dissembling the sheet back into individual frames.
With exactly 1,424 images to be uncensored, I think we still need a better process to come along.