It's been a long while since I contributed anything to this website. I stopped making cartoons on the internet for fun many years ago as I focussed on more commercial endeavors and slaved away on soulless animated children's TV shows which you've probably never heard of.
The Chinese plague has allowed me to reflect on my life and really examine and what is and has been important to me. One of those things is stupid cartoons made by a small community of internet dwellers.
Newgrounds has always been an huge influence on my personality, creativity and sense of humour. In my opinion, the legacy of this place is severely underrated. I started lurking on this website when I was 13 and started contributing not long after. Now I'm almost 30 and I regret that I've never been part of anything great.
My presence here died at least a decade ago but today I come back from the dead trained and with new skills.
Shut up and get to the point!
I want to start an old school Flash collab based on a crazy idea I got during a boring quarantine, sitting in my office drinking way too much coffee and chewing way too much nicotine gum.
This week - during my days off - I trained a neural network (artificial intelligence, if you prefer) with a corpus of transcripts from almost every single episode of the hit television program Friends.
The source .txt contains approximately 700,000 words and I threw a couple episodes of Breaking Bad in there for flavour. I then left it to train for a few hours, and again for a few more and then once more for good measure. The sample texts that the AI generates are interesting, to say the least.
What's the idea?
The idea has the potential to be really great or really retarded. Either way, I want to try.
My vision is a montage of short animated scenes (10 seconds to a couple of minutes) based on the TV show Friends using only the generated screenplays from the AI.
The results will be bizzare and avant garde but potentially hilarious.
What the hell do you mean "AI generated scripts"?
In February 2019, OpenAI released a paper describing GPT-2, a AI-based text-generation model based on the Transformer architecture and trained on massive amounts of text all around the internet. From a text-generation perspective, the included demos were very impressive: the text is coherent over a long horizon, and grammatical syntax and punctuation are near-perfect. The most popular AI text generator is TalkToTransformer. Play around with that to get an idea.
There's also been countless articles written about neural networks and how they're being used and these guys even made a movie with that nerd from Silicon Valley based on a screenplay written by a neural network.
For this, I trained my own model on my computer with a GPU following the tutorials by this guy. Here are some choice samples:
[Scene: Chandler and Joey's, Chandler is trying to find Rachel, Rachel is there.]
RACHEL: (starting to cry) Why'd you get me married here?
CHANDLER: Well, I thought you were probably gonna say it.
RACHEL: Well I think that's why I started asking you out on a date for Ross.
CHANDLER: All right then! (they start arguing.) Oh no, Chandler I just said you could sleep! (to Rachel) You can, we can just have fun, we can do the wedding! All right?! I will give it to you! (to Chandler) I love you!! (to the baby) Oh my God!! (to Chandler) Oh, you did not realize that I had just done y'know…
RACHEL: Yes, that's it's fine!! (she hugs him tight.)
JOEY: Come on you can go out with a real man! (to the baby) All right, we can all live together! Y'know? And I love you!
MONICA: I was gonna say it but you have been really great to me.
J.D. CHANDLER: Yeah, I'm not gonna go home and cry about the baby! (to the baby) Well, you're never gonna cry because of the baby!!
RACHEL: You're right, it's just not a very common occurrence.
JOEY: Yeah but maybe I just thought I'd say it when I thought you were the baby.
CHANDLER: What if it was a man?
As of right now, the AI generates a lot of banter about "the baby". Lots of characters have babies, or talk to babies, or are pregnant with babies. I'll keep training the model to see what happens. Here's a scene that made me laugh:
[Scene: Monica and Chandler’s apartment, they are entering through the door as Chandler enters.]
CHANDLER: (entering) Hey!
MONICA: Who’s that?! (They both stare at Chandler as he enters into his apartment.)
CHANDLER: Hey, hey, everybody!
ROSS: (entering) Hey, guys!
CHANDLER: Hey, guys!
RACHEL: (entering) Hey!
ROSS: Hey Ross!
JOEY: All right now.
It's important to note that these aren't random sentences copypasted from random episodes and spliced together. This is the result of an artificial brain learning the language of "Friends" based on a 3 MB .txt file and like 6 hours of training.
There's also a way to realistically clone voices using neural networks but I haven't been able to figure it out. So if some nerd reading this is smart enough to understand this GitHub and figure out how to clone voices, I would love to integrate that technology into this project. I tried it but I couldn't get the repo working properly.
...but until then we'll have to settle with regular old humans. Which brings me to, probably, the biggest potential hurdle for this project:
Good Voice Actors/Impersonators
Considering that the source material we're dealing with is already nonsensical enough, what's really going to sell this project is talented voice actors. Talented voice actors who understand the characters they're impersonating (their mannerisms) will really carry this project. The scripts are heavily dialogue driven, usually with just one scene (location) and some basic actions. Although there is lots of room for the animators to have fun with this, a solid radioplay is absolutely essential for the storyboard artists to play off of and really be creative. The VA's will also make or break whether or not the audience understands this as a "Friends parody" or just some weird animated sitcom.
I'm not entirely sure. I guess partially because there are a lot of episodes which also contain a lot of dialogue with specific speech patterns which made training the model easier. It could also be that the transcripts are readily available online in simple text format, which made compiling the corpus fairly easy. It's also one of the most popular TV sitcoms of all time so there's broad appeal for this subject in parody which could help with views on our content. I hate to piss on creativity with marketability but it helps to choose a character or subject from pop culture if you want people to watch your cartoons.
How's this going to work?
There's no need to be overly ambitious so we'll start off with one episode. One episode is one collection (or; montage) of disjointed/unrelated scenes which are animated by individual contributors.
If there is some serious interest in this project, I will go into detail about the pipeline and how the process will work but for now I'll just leave this as an open call for talent.
- Voice actors. (Very important).
- Board artists (can be animators)
- Designers (characters, backgrounds, props)
- Animators (can be board artists)
- Musicians for the theme song.
I will be directing, producing and editing the project but I will also do boards and animate some scenes. I can also possibly offer revisions to some of the junior animators although I would prefer this project to have a certain level of quality.
Who knows, maybe I'm going crazy in quarantine and this isn't that great of an idea. Either way, let me know what you think. If you're interested, post a link to your stuff and we can start from there. There is currently no limit for contributors.