How to make an AAA game in Unity (or fail badly)

Well those voxel data are born from a function, it's not different. The function render to voxel instead of pixel that's it. They did work on their function expressiveness though. You can't store 18 quintillion voxel data.

Anyway, there is plenty game that are much better than no man sky, it's pretty barbone, their art direction struck them a deal with people who gave them plenty coverage, they were lucky, that's all, you can go see the crying river of salt of those who did it before them ...

Also ps4, though I want to see what people do by ripping the game apart and modding it.

 

That's familiar, either you linked to it previously on the thread or I ran across it. Either way, definitely worth re-watching!

 

This is also rather cool if you try to do AAA level stuff with minimal work and cost

 

Another cool stuff from naughty dogs

paper here
http://advances.realtimerendering.com/other/2016/naughty_dog/index.html

They rally got full procedural whenever they could!

 

Hooo boy, here we go again with "minimal work and cost." Photogrammetry is a lot like shooting for film, and I think people don't get that, or they don't get how expensive that actually can get, especially since photogrammetry is much more finicky, especially for environment work. You basically need perfect weather conditions every time, you need to be very much aware of how to use your hardware or willing to hire somebody who is, you need to be able to tell at a glance what things you can and can not use (complex objects with lots of little details are a no-go), you need to be able to manage the workflow for handling a huge amount of photos (even if you're just doing it for a few props you're going to have thousands of photos), and this doesn't even begin to cover the cost of colour balancing everything manually (which you will have to do

Photogrammetry is a lot like 3D scanning in that it changes the workflow away from certain other methods but comes with its own very expensive caveats. It comes with a large upfront cost (lighting equipment and cameras aren't cheap, even to rent) and requires a lot of effort to properly stitch everything together.

 

Have you watch the video?

Also I have already played with photogrametry
A lot of the stuff you mention is less expensive than a skilled artist doing hi poly anyway. The video show some practice and pitfall, I know it's not magic (very specific type of material only).

1. For a lot of object you can do it in house with relatively minimal set up.
2. Best practice is already diffusing itself, it's not as hard as doing creative photography. Rules of thumb is not expensive. Waiting for good weather is not expensive.
3. Workflow to manage huge amount of photo? I mean seriously? I don't need to have photogrammetry to have this problem. You still have to hunt for references. If you do that, you will manage photogrammetry.
4. Manual labor is the most costly here and the point is that there is more and more documented workflow to made it automatic. Aside maybe from masking. It's currently a focus of many people, so expect solution very soon. Especially with a macbeth.
5. At least you don't mention dense mesh like other skeptics, it's not supposed to be used as is ...

Anyway the technique is more and more mature, there is a cost in learning the skill, but saying it's too expensive that's it's out of indie reach is not true. And now practice is well diffused, the cost of learing is going down. The vanishing of ethan carter was made using photogrammetry, it's an indie game with a very small team, they pioneered it. The cost is really going down overall.

It does have a cost (good equipment as you said, but a lot can be dodge with DIY, I would say a good camera is unavoidable and it's likely that if you are a professional that you will have a good computer too). But it's on the level of a single person too, which is the point of the thread.

I don't think it's productive to talk "skills", if you do AAA this should be implicit. The real problem is cost (time and money). It does matter, but it always does, even for procedural generation, even for basic modeling and texturing.

Personally I'm experimenting now with cheap camera, I think it might be possible to get something interesting by doing thing differently (My end game is to see if I can do something with a single board computer to automate capture). Don't really know yet, I have made some test that was not good but better than expected, while it's not currently planned to make new test in the present (I'm having different priority now) I hope to get back to it and dig this path.

 

There are places already selling this sort of stuff. Once a large enough collection is available, it's going to be one route to easy access, high quality data.

 

Photogrammetry of face (especially animation) is still very very expensive, true lol!

 

Seems like a bit of scale contrast was really needed, this video is more of a proof of concept than an actual balanced settings

 

Is Unity the right engine for this?[/QUOTE]

I just had a good laugh when I did read the last part of this post

and no man's sky.

But yeah, nobody could have known back them how it would end xD

 

No man's sky is still polished indie, it's impressive for a small team and certainly not laughed at by game developers. My little studio couldn't do it.

 

Just look at other game with space to planet seamless transition:
- space engineer
- dual universe
- can you name some other?

They have programmer art ie what people call progen is not bland
Artist driven procedural generation seems to get you further, though what nms lack is not better progen but more "functional" game design, visual variety without variety of meaning feel samey.

 

It's not quite seamless planet transition. It's close but no cigar, there's a visible transition if you watch out for it, but very understandable...

 

Do you mean these games or no man sky too?

 

I found something for your quest for the best afro hair again:

http://wiki.polycount.com/w/images/4/42/Fluffy-fur-breakdown.jpg
http://infinitasuperficies.com.br/apresentacao/diorama_ovelha.html

These were mentioned in the context of being a reference for making fluffy afro hair on the polycount slack chat.

 

Thanks!

I'm aware of the shell techniques, I'm worried it's not truly practical though (fillrates? transparency?) and it has some silouhette issues. Looking at techniques to build tree was useful too! Right now I'm looking more about lighting issues (the assumption for hair are generally false for afro hair, the reason why they are curly is that they have not a cylinder shapes but the cross section is a flat ellipse or a crescent! and the detangled strand as a kind of random torsion applied to them), physics issues and hair styling issues! I have learn you can break down the hair shape in the same manner than flat hair aka the overall volume, some stray strand that break it a bit and the fluffy rebellious bits (basically universal principle lol, the real lesson is that observation is key in making sense of it). I'm still obsesses with perfect silhouetting but I have yet to build the right "head dictionaries" for the different "styling vocabularies" lol. I'm also looking at stuff like moss shader to get ideas to simplify a potential ground truth model after the facts.

I'm still looking for a good tools that would allow me to paint anisotropic map (I'm not using photoshop so the polycount's hack won't matter), maybe I will have to code one? I'm not a coder!

I haven't actually apply my discoveries in practice, but I think I have now enough data to try and test all different approaches, I'll start with a throw away model to warm up and then planned a proof of concept character! I'll share my finding as I go along soon!

 

Take 2

It seems I wasn't the only one to make that observation...

However the game still have incredible landscape pretty consistently relative to other procedural game! too bad the game don't have a gameplay that take advantage of them and give them actual meaning. I have seen many stuff I though "this would have been hand crafted!". This give me hope that somedays someone implement actual gameplay function and better man made structure on it with circulation.

 

Space Engineers has two major differences from No Man's Sky though. It doesn't have procedural planets and the planets it does have are loaded entirely into memory from template files. Without planets you can play Space Engineers with about 8GB of memory but once you enable planets the actual memory requirement becomes 16GB (anything less crashes).

No Man's Sky approached it in an entirely different manner due to the limitations of the consoles. It generates the procedural data for the game on the fly as needed and once no longer needed it purges the data from memory. The pop-in effect we see in the game is entirely due to the processor being unable to generate the data fast enough to be seamless.

 

Yeah, I know that, I was talking about the visual mostly, most other game that looks like no man sky in concept, whatever different they have, just don't have the flair of nms art direction, they look blander in contrast.

 

That sounds pretty cool. Does it also mean a savegame of a planet with lots of mining and building done to it creates a several gigabyte big savegame? Some day I might need to give that game a second chance. On my first try I was overwhelmed by the complexity and the feeling "this feels a lot more like work than like playing a game".

 

I don't know. I only knew about the non-procedural nature of the planets from the reddit community.

 

IF they're loaded in their entirety as Ryiah said, then probably not, unless you BUILD something big.

I have that game in my library, but I haven't played it much for two reasons: it gives me motion sickness, and I disliked clumsily done building/vehicle editor.

 

This. I wish it had a system for building vehicles similar to Besiege where you leave first person but so many sandbox games seem to insist on being first person for laying out the vehicle. It just makes it so much harder.

 

Which reminds me I need to go back and play the new Besiege levels some time. I bought it rather close to launch and haven't touched it since after I had finished the levels they had at the time.

 

unfortunately this thread remind how hard the market is, even if your game look AAA, remind me an article i read recently that anyone should read befor starting game develepment
http://indiegames.com/2016/09/qa_arcen_games_chris_park_on_r.html#more

 

If you are a small studio, employing people, and go over budget.....Not everyone who develops games have the same goal. Good read but doesn't apply to most of the people here.

 

Perhaps applies if only from the perspective that a person should always stick to what they are best at. If these folks had great success from their AI War game... game proved to be both something people wanted and something they were good at developing in a cost-effective manner.... makes me wonder why they ever changed direction to begin with.

I think every Indie needs to realize that to succeed they will very likely need to be making game after game very similar to each other. Once they find their first sucess I mean. Until that point sure make all kinds of different games. But once you make a game that sells that is your path forward.

I guess the TLDR is simply: focus only on what you can do best (or at least focus only on what works... and then get better at doing it).

 

I would reword that to:

Only experiment and waste cash on that if you HAVE cash to waste.

Keep your cash cows alive and milk them regularly to build up the war chest for those wild experiments that can pay off dearly if they work, but will bomb often.

Just make sure when you are working your cash cow, to not overwork it to the point were fans are turned away from the series.


Problem is not trying something different. Everyone who has the capacity to do so should do that, not to get caught in a difficult spot as soon as the only product he sells is no longer in demand, or flooding the market with too many sameish products.
Problem is that obviously this company didn't work in a well planned, business like manner, at least that was how I read the article.

"We found that Raptor model on the Unity asset store and found its jerky movement funny. So we decided to build a game with that..."
Wait, what? Where is business plan? Where is the GDD? Where is even the high level vision, if that REALLY was all that sparked a new project that would have costed multiple 10k$ in the end?

I appreciate that small Indie shops do not work with the same rigid project structures bigger shops do, or do experimental / game jam like projects often.
Either this game should have been cut before reaching early access, or the development should have been structured so the early access failure was actually part of the plan in a "lean startup" kind of way (though you can only "market test" one or two games this way until the steam community catches up).

In the end, it was a wild, quite unplanned experiment from the sound of the article. It had a 50-50 chance of succeding at best and bombed. Hardly something extraordinary, when some big titles, carefully planned and well executed, are bombing in the market too.


Next time they try something different, maybe spend enough time on the planning and market testing side so they can explain more about the origin of the project than just an anecdote about how they found this cool model on the asset store.
Or go into the project knowing it might bomb and spend the minimum on it before going public with it (Kickstarter might help here).

 

After Gar mentioning that this is related to studio which made AI Wars, I've read it.

Nice material here:

 

Early access was design first to show step of development, now the expectation seems like it is a kind of beta with cut features, ie that it must have quality from the get go (as in quality asset and gameplay integration) which is not like dev friendly at all. I think a better strategy could to have prototype incubator shown as very small game, to look for concept that actually works from a very rough state with a much lower cost than the creeping asset inflation of current early access.

 

This is huge HUGE, voice might be in reach soon

https://deepmind.com/blog/wavenet-generative-model-raw-audio/

 

Well, really, what did you expect?

You are putting a game in front of the unrestricted public... Without all the caveats that come with Kickstarter (where you do NOT put a game in front of the public, merely a concept, maybe a demo).

To the general public this sounds extremly like an open Beta.


And to be honest, it would be quite STUPID to put your game into early access if you have not at least a very good vertical slice of the product. At least that part should be finished to a high standart, polished, and almost bug free.
I really like how "Ghost of a tale" handled the early access. Its the first few levels of the finished game, with all the polish expected from a finished games, and just some few bugs left.

THAT is how you do early access. You Shock and Awe the guys who spent on your game on early access with a good, if rather short compared to the full game, product. Something that makes them craving to get their hands on the full release. People will say "its not entirely finished, thus rather short, but DAMN, is that short expierience GOOD!".

Compare that to building the full game, but without polish and with many bugs left. People will stop playing after a short while because "Its an ugly looking turd full of bugs... yes, there is potential, but there are only so many gamebreaking bugs you can take, and still continue to stare at the ugly placeholder art!".


Now, if we are talking about NOT investing the time to build at least 20-30% of your final game to almost a final state when it comes to polish and being free of bugs, OR build the full game to a lower standart, but release the buggy prototype you built in a weekend of that awesome game idea you came up last week, and asking people to chip in 10$ so you can actually build a prototype complete enough to figure out if this concept actually works, I have little sympathy left for you.
"Dev friendly" doesn't mean "Not taking risks, not chipping in a good portion of your own money before expecting others to pay anything"... "Dev friendly" to me means "Open to anyone who has a solid plan and game, is ready to take some risks and has the needed budget and time to build the initial offering".


If "Dev friendly" turns into "Customer unfriendly", your early access customers will start to loose interest and just ignore early access titles altogether.

 

I think a diagonal slice is better, early access is to find early adopter, game like unturned, minecraft, totally realistic battle simulation or gang beast thrive with minimal graphics to focus on having good control and gameplay. The exact opposite of release raptor who have somewhat polish graphics and gameplay isn't there yet ... the core should be there whatever you do, and I say diagonal slice because some people really overload the word MVP

 

I found a video of bioware about procedural cinematics, that was teh early stuff but it talk how it works, they perfect it with swtor (almost entirely procedural with text input to drive the generation through emote tag).
http://gdcvault.com/play/1017703/Directed-Generation-Procedural-and-Placeholder

 

@neoshaman great stuff, thanks for the link. At this point, I could imagine it being possible to procedurally animate (or populate with animations) cutscenes or dialogue interactionsn in a production indie game but voice-over remains the huge potential cost. I've had a look at a bunch of software for text to speech but none of them sound anywhere near game-ready. The idea of slicing up recorded speech sounds interesting, I'll have to have a look at it in a bit more detail.

 

I think Text To Speech technology will fall very soon in the next years to a surprisingly useful state (if you have click on that google research the video is talking about) thanks to the discovery of wavenet. By that I mean generation of intonation and transfer of style from voice to voice. Still won't be AAA level (in that timeframe, but who know in 5 years) but might be decent enough so that quantity trump quality (ie everything voiced). Also it's promising for music, current test are not guided, so who know what it can do if guided information like with the voice. The thing is that it's not a new technique at all, it's the same technique used for image generation that have rocket shoot past most optimist predication (it's production ready with very convincing product on the market see prisma), it will benefit all optimizations.

https://deepmind.com/blog/wavenet-generative-model-raw-audio/
It's not ready as in it's not ready made and not real time (which is the end goal) and there is style transfer experiment yet. But you have enough information to create one yourself.

The thing is that good enough sub AAA is already achievable, so that's a win.

 

Yeah I saw that, it's by far the best one I've seen and doesn't glitch hard like a lot of them do, but the tone still sounds fairly flat and much like a robotic call operator. I wonder if it can be tweaked to use a more dramatic way of talking without the quality breaking down?

 

It's a neural network, trained, rather plainly, to make simple correspondence between words and sound. I guess the exact same set up but with emotion and intonation annotation might have some result. Now where do we find the annotated data? I'm sure some game or soft have library of sound with variation of the same thing with different annotation, if we could hack some
You have to understand that it works based on a very generic method, not at all adapted to sound to begin with (it's a plain convolutional network). Neural networks are magic again.

 

Well the paper describes two different ways that the neural networks were trained to produce different kinds of outputs:

a) Fed a lot of recordings of human voices, to produce a non-existent language but with realistic (more human-like, personal) intonation;

b) Fed a lot of text, to produce coherent (english or mandarin) speech - but more robotic sounding.

It's hard to tell in what ways if any these could be combined. The made-up language sounded a lot more emotional and 'alive' whereas the text-to-speech sounded robotic. If they could be combined to produce a more personal, warm or emotional text-to-speech that would be very interesting.

Another thing is it seems at the moment that the main problem with the text-to-speech is the data 'filter size' in that the neural network can't build up to emphasis over a long period of time - it forgets relatively quickly the previous data and can't use long-term data (more than a second or two) to construct a longer-term structure such as, e.g., building up to a dramatic delivery. This is the problem with a lot of text to speech that I've seen - there are no low-frequency structures in it and each 1-2 second fragment sounds similar. But these low-frequency structures are very essential for dramatic speech.

I wonder if non-realtime procedural construction of speech (e.g., built on level load) would be possible to enable a larger filter to be used and better quality achieved.?

 

I've had some ideas where the current state of TTS would be adequate, but every TTS option I can find has too many limitations on usage. Heck the built-in stuff in Windows and Android would be fine but both companies restrict how it can be used. (I don't remember the details now, but I'm not working on that idea, and that's the reason...)

 

That's the current area of research about "temporal" neural network now, they have short term memory, it's an active field of research. That's what I'm hinting when I say it's a generic technique, the solution might be in other papers that look for other type of data, like the current super resolution craze (bringing small image to a bigger clean resolution). It's an active area of research and past the jargon that's essentially different way to connect the NN together. If the memory improve, it won't be only for TTS, generated text will make more sense too, so on average you won't just be generating sound, you would also be able to generate meaningful text with it at the same time, basically artificial speech.

About practical use, the way it works with "robotic voices" is that it simply map sounds pattern to a string of text, for the NN it's just random statistical noises to understand. The voice is robotic because it found a canonical sound to the text input for generation. Now if you were to feed annotated text (aka text + intonation information) he would discriminate between the different intonations. Now you would feed annotated text and the NN would rebuild a voice with the intonation, memory wouldn't be a problem because the input would be the low frequency structure and has all the necessary data! Maybe the only adjustment outside of input data would be adjusting the hyper parameters of the network, but that's to test. Also I'm not sure how fine the control would be too, finding a system of annotation that make sense in an objective way even to human might be the main problem, it wouldn't be the NN per see.

I would be fine with a current TTS solution too! for some cartoon game like tomodachi that would be perfect, being able to procedurally generate voices with their own personality.

 

I looked for free and open TTS solution
I found that
https://www.efrontlearning.com/blog...ource-text-to-speech-tools-for-elearning.html

 

Maybe a more useful (and fun!) method would be for you to read the text to the software, and it extracts some basic intonation data (not enough to characterize the voice but enough to give it meaningful structure) and then reconstructs the text into speech using that intonation data.

 

You mean mapping voice to voice instead of text to voice? So basically it would map your input voice to target voices? You would need pair of your voices/actor voices with the same intonation.

That said I was wondering, can we already sensibly modify a regular voice characteristic artificially? I mean if the system only generate one type of voice and you modify it the usual way to create new voices through filter (pitching up and down or autotunes lol for the more common)! It wouldn't be AAA but might work in some cartoonish way?

edit: Okay lol I misread, you mean you are the actor that feed the data! yeah of course lol post process voice modification is still something to look at (I'm no expert!)

 

Yeah, the idea is that I don't need to be a good voice artist, I just need to give it enough intonation data to form a structure for its own processing work. I'm sure that in anyone's voice there is a lot of retargetable structural information that could be useful. If the synthesizer could use it to come up with a range of different voices that would be great.

 

You still need those annotation, else it will collapse in a statistical ideal voice per text, the reason the random one is more lively is because there is no discrimination with text toward the ideal voice.

The random one is just a predictive sequence based on what data came first, so the sound depend statistically on history. It's basically a better markov process. This won't be useful for actual voice generation.

What you want is that he is able to predict:
1. what sound pattern belong to text structure
2. history of intonation within emotional context
3. mix both

The problem is 3, basically it's 1 with context from 2, you would need to pass those information to the network, that is the same word have different sound pattern (the intonation) depending on history and emotional context.

Just create data as in: record text in different known intonation, reward the network for both generating the correct string of text and string of intonation annotation, then use it to generate back speech by inputting text and annotation. The system would be tied to your system to annotate intonation.

I term of data, you would need that it correctly output the right text from your voice with extra tonal information. Maybe the easy way is to train two network first, one for text recognition and one of intonation recognition, then use that to train a third network with data automatically tagged by the two first. That would generalize and automate the process to any sample pair of text and voice and guess annotation from them to feed the intonation networks. That said I wouldn't just use the automation, it might fail in many context (jargon, slang, social cues) so it's a very specific set of data you can have. Game though are a good target to rip data, they often have pair of quality sound with text, it only need annotation of intonation.

Also beware of overfitting, might not be a problem if your domain is exclusively your voice, but if it works with your voice, it might not works with other voices, you need to train with variety of data for good generalization. That may or may not be a problem since your problem domain might be small and we are using it for practical purpose, so as long as it works within the condition you design it to work.

 

I just read op, I also thought a lot about what game indie can, should make.

Of course, at graphic and programming, contents quantity, indie can't compete with AAA, not even close.

So only hope is story and game design. (game's 4 factors - story, game design, graphic, programming)

and your last presumption is hardly possible. One man, indie can't maintain productive will and passion for years to one game unless its alpha or beta released and already got much attention and money.

And you said many asset store's tools. Yes it is there. But those tool also required learning time, actual practice time to fusion with my game. If stuck occurs, contact with that tool's creator,,, etc... whole this will also consume vast time.

So narrowing scope down of making is only possible, of course at the same time gamer's experience should be satisfied from that scope.

So here is dilemma of indie game and why it is hard.

 

Nan man, you are just not a samourai, but plenty people slave off for their art even if they are starvin'.

But most importantly it's all about reception and illusion. I got a lot of flak (and rightfully so) in the lost soul inside thread for comparing it to the final fantasy xv downgrade. The reality is that I was talking about reception not production, lost souls was hiding its weakness smartly, you can have signifier of hi quality works without indulging into the wealth of production of AAA games.

For example AAA game would likely to have all their model uniquely handcrafted and correct all along production which is wasteful, indie could design a few character with part smartly reused (nobody generally checked that all your hands are unique) but keeping focus on element that attract attention (face is generally a focus).

AAA design to their whim, while an indie would have to design to take advantage of what signifier is easily available to strategically create the illusion of higher production, he would design with what's possible in mind, avoiding case that aren't possible, basically designing to play with the strength of these constrains.

For example if making too many human character is expensive, maybe considering creating a world where sentient robot and alien is there too, they are cheaper to model and animate because they can basically be anything, no uncanny valley, if natural cloth are add to do, go into the future, you can create crazy cloth that bypass any comparison to reality. Focus on a few human characters, with maybe some very few highlight complex scenes, design to gave a feeling of a higher bar, despite their small number, they will create a halo of quality that will contaminate the perception of the remaining aspect of the world.

That's actually an old advice, directed attention is very powerful to suggest more than actually exist.

Edit:
A famous example is star trek teleportation, it happen because they didn't have the budget for spaceship landing, it became like a signature things that had contributed to the identity of the show. It's not questioned because it totally make sense, it didn't appear as a cheap hacks it really was.

 

It's possible, but of course, why not release alpha and beta and get the train rolling early? Point is, if you can't get any good reaction with an alpha or beta you're probably wasting your time.

I believe that the only asset store tools anyone should use in this sort of situation are ones that they not only have full access to code but also have modified to suit their own purpose. In fact I think it's a good idea to spend some time releasing smaller games in order to iterate on the whole workflow that is going to be used.

Non asset store, well-established tools are another matter, there's a small risk of it being abandoned but it's just part of doing business.

 

Anyway, I think indie should make game like Tetris or Flappy Bird. It is perfect example of indie's successful game. Simple but has power of catch same-era's gamer's attention.

These are realistic scope to make for indie.