Jump to content

Sirus

Fuwakai
  • Posts

    4
  • Joined

  • Last visited

  • Days Won

    1

Sirus last won the day on September 1 2023

Sirus had the most liked content!

1 Follower

About Sirus

Profile Information

  • Japanese language
    Intermediate
  • Development skill
    Programmer

Recent Profile Visitors

The recent visitors block is disabled and is not being shown to other users.

Sirus's Achievements

Fuwa Novice

Fuwa Novice (2/11)

5

Reputation

  1. General thoughts If machine learning research keeps going at the current pace, AI will replace most jobs that don’t require robotics, I think this could come in less than a decade. Society will have to change and adapt, something like UBI (Universal Basic Income) will have to be put in place or things will be dire. I think this will happen whether we want it or not. The question is whether this will stay locked behind big companies controlling everything like OpenAI, digging the gap between the classes even more, or the revolution will be open sourced and the power will be in the hand of everyone (StabilityAI, Meta, all the smaller open source projects…). It is my belief that creators should start getting familiar with these tools starting today and include them in their workflow to boost their productivity, especially if they’re working in a commercial fashion, or they are at a higher risk of being replaced first by people with higher productivity due to these new tools. I fully understand the passion people have for their arts, but if it’s your livelihood, capitalism will have no pity for you, and you will have to follow the market to hope survive. One of the most positive points for me is that it will usher a new area of creation for indies who work, like me, solo, and don't have enough lifetimes to learn all the different skills nor a bottomless piggy bank. I think we'll see more indie games of higher quality popup instead of the nth pixel art, 8bit music that's too common. On the legality From all I’ve read, these models are currently legal in the U.S. and E.U. as of now. This might change in the future and might be cause for concerns, but hopefully things will be cleared up soon. The worst that could happen are draconian rules which would make it easy to comply for megacorporations and impossible for small indies/open source. On the ethicality The ethical debate has been mainly centered on the art side since it’s one of those that’s the most easily distinguishable, I imagine with voice & music models getting way better it will spark it up again. I personally don’t think it’s immoral to train on copyrighted assets of others AS LONG as you’re not purposefully trying to copy their style. Models such as stable diffusion use billions of images and so the weight of each artist becomes very diluted (but some bigger artists can still be recognized thanks to the amount of art they have available). No one should train on someone’s creation with the goal of avoiding paying them. I think an exception is copying the style of a large production to make non-commercial fan content (see the recent debate with the Scooby-Doo controversy). I will now go over the different domain that applies the most to us as game creators. Art models This is clearly what sparked the whole debate in the recent months. In a year since the release of Stable Diffusion 1.0, models have made giant leaps (SDXL, Midjourney V5…) and have an arsenal of tools (ControlNet…) making them some cases almost indistinguishable from human creations, be it drawn art of photography. While they can deliver stunning artwork, I think they still very much struggle to make anything consistent, which is a problem especially for something like a VN. I think these tools are best used by an already competent artist, as they could draw a sketch, feed it to ControlNet which keeps the consistency of what’s drawn, inpaint to get some of the detail rights, and finish up the fixes in photoshop. Can you make a full game with AI art without any artistic skills? In my opinion, no. Not something that will look good or have any kind of artistic direction anyway. Code models I will be biased here as I’m a developer by trade. Code models (and by extensions more fully fledged LLM like ChatGPT) have taken the development world by storm, claiming you don’t need a developer anymore or whatever. That’s simply false, but not totally. At this current stage, they’re extremely effective as speeding up the workflow of a seasoned developer (GitHub reports up to 30% which is about what I would have said myself using it daily https://github.blog/2022-07-14-research-how-github-copilot-helps-improve-developer-productivity/), which could mean, for example, that a company could need one less junior developer to go with a senior developer, but that remains to be seen. The effectiveness is also pretty dependent on the language used, with something simpler like C# or JS working pretty good, and something like Rust struggling more. I use it myself as more of an autocomplete at a single line, which it excels at. It’s also amazing at repetitive code with slightly changes, which it’ll suggest instantly. What it’s really bad at, is following long list of instructions, or trying to implement algorithms that are not already seen hundreds of times on StackOverflow or GitHub, and this is where it stops being useful to a novice, and why I don’t think the programmer job is gonna disappear anytime soon (nocode platforms haven’t killed us after all, and WordPress might be a huge chunk of the internet but it doesn’t stop people from recruiting more developers than those that exist). I think this video is an excellent example, a man with no programming experience trying to make a simple platformer using only AI. Sure, he kinda managed to get something, but it took him 8 hours (or 4, I don’t really remember), when it could be whipped up in under an hour by a professional programmer (or probably even less than 10min without the assets). https://youtu.be/IyKKhxYJ4U4 There’s also a lot of these videos on twitter et al, of people whipping up a whole website in a single prompt with ChatGPT. It might work for some simple ideas, but good luck growing it up further. If all you have is a static page with some text then I guess it’ll work, but the average person won’t even know how to host a website. Text models ChatGPT also created quite a lot of turmoil, for good and bad reasons. I’ll mainly focus on 3 parts that could be relevant to us: Marketing, Creative Writing & Translation. For marketing, I think it can be excellent, especially if you don’t like writing overly formal stuff. It can be a bit dry, but I think it’s wonderful if you need to generate a lot of PR material as indies are often pretty bad at that. I think this usage will grow more and more. For creative writing… Well, it’s not good. It will write the most boring and generic stories, and the small context makes it impossible to keep track of all the elements for longer period of times leading to all kind of inconsistencies. I think having ChatGPT write your whole game would clearly be a mistake. Now there can be some more local uses. I like to use it sometimes to suggest different ways of phrasing or reformulating a thing (and I pick & mix what I like the best, if anything). It probably can be used for some more generic descriptions (that would be outside the narrative, like some menus or something). Also brainstorming some ideas, splitting tasks into some smaller tasks (that I do sometimes, and it suggests some things I could have forgotten to do). But please, don’t use it for creative writing, no one will want to read it. For translation, there’s some good and some bad. First off, the good, is that it’s generally better than DeepL and Google Translate. You can feed it some context, it has a better understanding of the world and what the specific word mean in this context, you can nudge it with some additional infos (especially the name if you’re translating to/from Chinese or Japanese). Now the bad… It’s better, but nowhere near perfect, will trip up a lot, get the context or the subject wrong (especially with Japanese where it’s implicit a lot of time), completely miss jokes or puns… The list is long. If you were planning to hire a TL for a cents per word then I guess that it can’t be worse, but otherwise, please do your international players a favor and hire a professional translator. Music & sound models Honestly not much to say here, I don’t follow those too much, but they seem to be making some progress. Sound models seem to be able to generate some common sounds quite well (barking, quacking…) albeit with some random noise sometimes, but then again there’s already giant sound banks that exists so I’m not sure how useful that will be. Music models seem to be making some good progress as only a few months ago they sounded like some melody straight out of hell, and now its’ almost listenable garbage. But they will keep making progress, so we’ll see. Voice models This one has caused some controversies recently, as people have claimed their voice was used for training. The TTS models, which have been a thing since forever but only recently lost most of it’s roboticness are becoming straight up amazing, and something out of ElevenLabs which seems to be the SOTA (State of the Art) right now are almost indistinguishable from a real human voice if you add a little background music. Though the emotions and range of voices are still narrower and it clearly won’t replace the capabilities of a real VA for a more complex job right now, it could be exciting for something like a RPG with hundreds or thousands of characters with their own unique voice. There’s actually multiple parts all making progress at the same time, there’s the voice generation where you can generate new voices without any input, which could algorithmically offer an infinite amount of voices. Then there’s the clearly most controversial one, the voice cloning, where you can recreate pretty accurately the voice of someone else from an audio clip as short as a few seconds (but the more audio and the higher quality it is, the best the results will be). This can lead to funny clips of Trump and Biden talking to each other, but more nefarious use like fake news or cloning the voice of a VA without their consent. This last usage must be firmly condemned, and might even be illegal from my short researches (https://higgslaw.com/celebrities-sue-over-unauthorized-use-of-identity/, but I’m not a lawyer). The last part is voice morphing. You might have heard of it recently, with those videos where someone sings the song of another artist. This is a bit different from voice cloning, the cloning part still exists of course, but you’re adapting the voice to a preexisting audio which can be damn convincing. I don’t really see any usage in video games for that last part since you would already need a competent singer to do the initial singing which is arguably the hardest part. I think these models are gonna keep getting perfected and will lead to a lot less demand for VAs (except at the top, those who can sell a product by their voice alone). I can easily imagine some VAs, instead of coming in the recording booth, licensing their voice for a yearly usage, or by number of words etc. I hope I didn't miss anything important, and I'm open to discuss with anyone with an open mind
  2. I'm not sure I get your idea. I'm looking at writing prompt generator and I can find all kind of theme, from romance to horror. Why doesn't they fit a VN in your opinion? Same for the asset part, would it automatically create scenes using those assets for you? That's one of the ideas I had actually, but I wasn't sure because itch is already so big in the scene. What kinds of feature do you think itch is missing specifically to fit this role? Or is it just because finished games are mixed with unfinished ones, so the unfinished ones have lower visibility? This could be a really interesting idea to develop further, but I need to understand what the added value would be over something like itch, as it would be impossible to compete on the same feature set since the userbase is already so big.
  3. Hello, I'm a developer looking for ideas for tools I could work on that could fill a hole in the VN ecosystem. Has there ever been any tool you'd wish you had while working on your VN, or even before starting work on it, or when it's finished? It might be anything from a software, a script, a website (ex: a site to find other team members, a script to check grammar, a software to outline your scenario, etc.), for artists, programmers, testers, or any member of the team. I'm interested in hearing your ideas as detailed as you can, what use case would it fill, how much time it would save you, etc.
×
×
  • Create New...