Discussion AI Generated Videos
- Thread starter mrmaninblack2
- Start date
-
Welcome to the Fakes / AI / Deepfakes category
When posting AI generated content please include prompts used where possible.
How to inpaint videos using EbSynth:
EbSynth download here:
1. Save a video as individual frames in a folder, make sure each frame image is numbered (this can be done with Photoshop)
2. Inpaint over any of the frames in the video and save to a seperate folder, name the inpainted frame with the same number as the frame from the video (usually only one frame needs to be inpainted for one shot)
3. Open EbSynth and set the “Video:” section as the folder with the original frames, and the “Keyframes:” section as the folder with the inpainted frames. Click “Synth” and the frames for the inpainted video will be downloaded into a folder.
4. Put the new frames together to make a video. This can be done with Premiere Pro
EbSynth download here:
1. Save a video as individual frames in a folder, make sure each frame image is numbered (this can be done with Photoshop)
2. Inpaint over any of the frames in the video and save to a seperate folder, name the inpainted frame with the same number as the frame from the video (usually only one frame needs to be inpainted for one shot)
3. Open EbSynth and set the “Video:” section as the folder with the original frames, and the “Keyframes:” section as the folder with the inpainted frames. Click “Synth” and the frames for the inpainted video will be downloaded into a folder.
4. Put the new frames together to make a video. This can be done with Premiere Pro
Last edited:
Mostly no, but i'm a perfectionist. In fact, as it currently stands the state of AI generated (human impersonation) video content currently is (at best) a(n arguably) worse version of deepfake tech. While it is quicker currently, a good AI 'artist' will have almost all their results result in massive flicker issues, even using the best samplers/prompts/GPU's/RNG/ETC. A lot of these issues can be diminished using multiple techniques and having knowledge and ownership of different video and photo editing software, but at that point you're mostly just editing individual frames one at a time which is barely an improvement on itself.
Our biggest issues are:
1) text2video prompts come out a garbled mess, like looking into someone's nightmares.
2) video2video is decent but leaves flickering issues even on almost-perfect settings
3) Every single body part that moves needs to be measured and accounted for precisely otherwise you get flickering
Cartoon2realvideo is barely achievable currently, real2cartoon is probably the easiest, and real2real comes out with the closest results that you're looking for but with the aforementioned problems.
I've been working on different techniques almost everyday for over a month and my results are not up to my standards which is why I haven't released anything for a while.
Per examples:
Tl;Dr: We're not there yet. We're close, but we need better extensions and methods. The best we can do is faceswap videos to give a (((SEMI))) convincing output.
Can you explain cartoon2real and real2cartoon? What software do you use and is there any examples
WebUI Automatic1111. It's just using img2img.
The issue is IRL videos have more moving parts than a cartoon making smearing inevitable. To counter this you'll see AI video apps (only paid subscription ones for now) use cross fade blending to transition during scenes once it detects smearing. Unfortunately there's not really a way around this due to something not being drawn that suddenly comes into frame that EBSynth doesn't recognize.
Things EBsynth and AI have issues with are shadows, black colors, white colors, reflective fabrics, and straight line geometry (lines, triangles, etc.).
So while cartoons are easier to draw over for AI and EBSynth, making a coherent AI deepfake video is (to put it bluntly) a fucking nightmare no matter what programs you use currently.
What alleviates this a bit is using controlnet and playing with the values for HED, normal, depth, and sometimes canny and scribble. However at 60 fps which accounts for 1440 frames a minute with a typical error threshold rounding out at 50-150, you can easily spend all day just working on a minute long scene.
You could also mitigate this by lowering the frame rate then upscaling in Topaz or something, but it can sometimes make the flow feel more unnatural that way.
Taking this video for example, was just one frame edited and it took me close to 3 hours playing with the settings at 8gb vram using SD 1.5 as a base to get it barely coherent and I still consider it a piece of garbage.
ok but is their an free alternative for combining the videos, i use clideo but it wont let me restore the original fps only get close to it
The new TemporalKit works really well. Tutorials and extension can be found However you NEED to have decent skills with controlnet.
What do you guys think of the work Runway Research is doing? They're calling it Gen-2, capable of transforming text to video. The concept isn't anything new of course, but it's honestly so exciting especially knowing how rapidly AI is developing.
The field is developing fast, and a hopeful part of me envisions a world where I can submit a model (Like a random eChick) and easily create video that is both realistic and natural, using nothing but text prompts and my imagination, but I know it will be a LOOOONG time before such things becomes readily available to suit our degen needs, especially with the active attempts of companies to block users from using their technology to generate NSFW or malicious outputs. But there's some pretty god damn smart perverts out there, and there's always a way to bypass restrictions ;)
The field is developing fast, and a hopeful part of me envisions a world where I can submit a model (Like a random eChick) and easily create video that is both realistic and natural, using nothing but text prompts and my imagination, but I know it will be a LOOOONG time before such things becomes readily available to suit our degen needs, especially with the active attempts of companies to block users from using their technology to generate NSFW or malicious outputs. But there's some pretty god damn smart perverts out there, and there's always a way to bypass restrictions ;)
Last edited:
It's good to be alive now.
I can foresee a future (maybe within the next 5-10 years) where we can generate our own movies, our own TV shows, our own music, our own books, and our own news. No need for authors, musicians, actors, producers etc.
Many of these possibilities are already available - at least in a rudimentary form. But to have an AI-generated movie in 4K with perfect quality - that will take a bit longer.
We will have sites like Onlyfans and Pornhub with AI-generated material.
Whatever you have dreamt of watching, whatever your kink or fetish might be, you will be able to watch it online or create it yourself - in the privacy of your own home.
I can foresee a future (maybe within the next 5-10 years) where we can generate our own movies, our own TV shows, our own music, our own books, and our own news. No need for authors, musicians, actors, producers etc.
Many of these possibilities are already available - at least in a rudimentary form. But to have an AI-generated movie in 4K with perfect quality - that will take a bit longer.
We will have sites like Onlyfans and Pornhub with AI-generated material.
Whatever you have dreamt of watching, whatever your kink or fetish might be, you will be able to watch it online or create it yourself - in the privacy of your own home.
You had me until "in the privacy of your own home." In 5-10 years privacy will be a memory.
I politely disagree. In 5-10 years I should be able to purchase a NVIDIA V100 Tensor Core GPU - or similar - and with the tools available by then, I expect to be able to produce all the content I wish - in the privacy of my own secure home lab.