AuthorTopic: NEW CLUSTER STUDY THREAD!  (Read 80591 times)

Offline Helm

  • Moderator
  • 0110
  • *
  • Posts: 5159
  • Karma: +0/-0
    • View Profile
    • Asides-Bsides

Re: NEW CLUSTER STUDY THREAD!

Reply #140 on: April 18, 2014, 03:06:18 am
Nope, there's too much parallax to check every frame. This is 900 individual frames for a 30 second loop. Thanks for the kind words, btw :)

Offline PixelPiledriver

  • 0011
  • **
  • Posts: 997
  • Karma: +6/-0
  • Yo!
    • View Profile
    • My Blog

Re: NEW CLUSTER STUDY THREAD!

Reply #141 on: April 18, 2014, 03:47:04 am
Thats really cool Helm.  :crazy:
The stars jump every 4th/5th frame to reach their destination.
It might help to do the math for a smoother movement.
And knowing that it is, we seek what it is... ~ Aristotle, Posterior Analytics, Chapter 1

Offline Helm

  • Moderator
  • 0110
  • *
  • Posts: 5159
  • Karma: +0/-0
    • View Profile
    • Asides-Bsides

Re: NEW CLUSTER STUDY THREAD!

Reply #142 on: April 18, 2014, 04:51:12 am
m-m-- math?!

Offline Joe

  • 0010
  • *
  • Posts: 298
  • Karma: +2/-0
    • View Profile

Re: NEW CLUSTER STUDY THREAD!

Reply #143 on: January 18, 2015, 02:57:39 am
Fantastic thread. I've been looking at it for a long time, today I finally set out to try it.  I feel the most appropriate name for this process is defragmentation.



To the best of my ability, there is no banding or orphan pixels. If there are, I was sloppy and missed it. I used 45 relationships sparingly and usually as a "cheat" to adhere to the restriction but still single pixel AA.

I found moreso than ever, I was catering to the pixels themselves rather than the drawing. That is, it's not as accurate as a painting would be, because (my own personal inaccuracy aside) instead of representing the source faithfully I had to bend to the will of the clusters.

Particularly difficult was avoiding banding. In a few cases (example: right eyelid) it was pretty much impossible to avoid; I had pixelled myself into a corner and the only choice was which two tones to band. The combination of no banding, no dithering, and no single pixels was very challenging. I will continue to experiment with this idea since one piece is quite insufficient, but nonetheless it gave me many questions.

1. Banding is probably the most incompatible with defragmentation. In many cases I had to contrive the subject, and especially the details, to avoid it and adhere to the restriction. My question: which, in your mind, takes precedence between the three? Banding, Defragmentation, or Fidelity? I have a feeling you'd say it's on a case by case basis. My personal tendency is to prioritize banding-elimination.

2. Why does this work? What is it about the absence of orphan clusters that makes good pixel art?

3. AA was discussed earlier in the thread, as something that is less necessary. Wouldn't sharper screens make AA more necessary than before? Or did I misunderstand that point

Quote
I'd have to practice a lot to see what that means intuitively for my own art

4. After a year of practice, where do the single and double pixel stand in your mind? As I understand it this exercise is meant to strengthen the integrity of your clusters, so that the single pixel becomes a tool, not a crutch. But there were many places where I knew single pixel AA would be the best solution, I merely refrained. I'm sure you don't think single pixels should be abolished; where do you feel they're appropriate?

E: Found more single pixels.
« Last Edit: January 18, 2015, 02:59:39 am by Joe »

Offline Ai

  • 0100
  • ***
  • Posts: 1057
  • Karma: +2/-0
  • finti
    • http://pixeljoint.com/pixels/profile.asp?id=1996
    • finticemo
    • View Profile

Re: NEW CLUSTER STUDY THREAD!

Reply #144 on: January 19, 2015, 11:30:26 am
Fantastic thread. I've been looking at it for a long time, today I finally set out to try it.  I feel the most appropriate name for this process is defragmentation.
That's an excellent name for the pixel art version of this.
For more CG-ish stuff, there are filters like GMIC's 'anisotropic smoothing', whose results, IMO, express the artistic ideal behind defragmentation -- everything being 'well fitted' to everything around it, and also show how not to apply it / when to break the rules (fine details becoming too indistinct). I mention that because they use a term in that that is also very fitting for the process we're discussing here, 'regularization'. It is more general than this specific process but captures the reasons for applying it.

Quote
I found moreso than ever, I was catering to the pixels themselves rather than the drawing. That is, it's not as accurate as a painting would be, because (my own personal inaccuracy aside) instead of representing the source faithfully I had to bend to the will of the clusters.
Of course, you could still bend more. Arguably the logical extension of these principles is to flatten planes in order to minimize cluster contention, ending up with something that is quite vector-y, which perhaps you could have used on the lips; I perceive them as more detailed than everything else.

Some of the stuff in ptoing's PixelJoint gallery demonstrates other ways of taking it further -- like exaggerating planes and faceting so that they are easier to fit together.


Quote
1. Banding is probably the most incompatible with defragmentation. In many cases I had to contrive the subject, and especially the details, to avoid it and adhere to the restriction. My question: which, in your mind, takes precedence between the three? Banding, Defragmentation, or Fidelity? I have a feeling you'd say it's on a case by case basis. My personal tendency is to prioritize banding-elimination.
I have the same tendency, but I am willing to confidently state that Defragmentation > Banding, due to mentally ranking them by how involved they are in making the overall work hang together. Fidelity is really a different type of thing IMO, fidelity is a rather selective thing in good art IMO -- you decide what aspects are important to capture and then do so. Fidelity to aspects not on that list is mostly a distraction until you reach the stages of seriously polishing the work.

Quote
2. Why does this work? What is it about the absence of orphan clusters that makes good pixel art?
Personally I believe 'regularization' addresses this question.

Quote
3. AA was discussed earlier in the thread, as something that is less necessary. Wouldn't sharper screens make AA more necessary than before? Or did I misunderstand that point

As I understand it this is mainly a question of PPI. The more pixels-per-inch our screens have, the less we need to use kludges like AA (sacrificing colors for the impression of greater resolution). With sharper screens, the pixels in AA are more apparent, but with higher-pixel-density screen, each individual pixel is less apparent (so, AA pixels are both less obvious and less effective). When pixel art was really big, we had PPI as low at 30-50. Now, 72 PPI is about the minimum you will see, 90-96 is commonest, and we are looking at even higher densities.

Retina displays probably need separate consideration, but what I say holds for conventional display devices (phones, tablets, laptops, desktop computers)

Quote
E: Found more single pixels.

If you use GrafX2, I made a script that detects them earlier in the thread.
If you insist on being pessimistic about your own abilities, consider also being pessimistic about the accuracy of that pessimistic judgement.

Offline Joe

  • 0010
  • *
  • Posts: 298
  • Karma: +2/-0
    • View Profile

Re: NEW CLUSTER STUDY THREAD!

Reply #145 on: January 19, 2015, 05:46:29 pm
Thanks for the well-thought reply Ai.

I'm sure there are better names to describe it, I think you got really close with 'noise reduction,' cause that's what it is essentially. Moreso probably than defragmentation—I was thinking of how all the stray pixels were being absorbed into their parent clusters, kind of like how data gets reorganized, but that isn't quite what's going on. I think noise reduction is a more accurate description and yes, this is definitely in the same vein as anisotropic smoothing. Culling and truncating also came to mind. I am aware of the caution toward inventing terms here, just trying to summarize this idea in my own mind, as well as distill it into a word.

Quote
It is more general than this specific process but captures the reasons for applying it.

I looked up regularization and found

Quote
In simple terms, regularization is tuning or selecting the preferred level of model complexity so your models are better at predicting (generalizing). If you don't do this your models may be too complex and overfit or too simple and underfit, either way giving poor predictions.

So if I'm getting this right, this is a less complex model, because it doesn't account for information that single clusters would. Or is it more complex because it has more rules?

Quote
Personally I believe 'regularization' addresses this question.

I tried to make this connection, but I don't understand it well enough yet. How does regularization explain why orphan elimination makes clusters stronger?

Yes you're right, fidelity is selective. I guess coming from a traditional background I'm so used to it being the point, but I'm starting to see how it's not the point with pixels. What you're saying is defragmenting is even more the point than banding elimination, that banding is almost accessory. I think I can see that.

Quote
Arguably the logical extension of these principles is to flatten planes in order to minimize cluster contention

Yes, if you were to take it a step further it would then be a matter of reducing the clusters themselves (like with the lips). Interesting... this makes a sort of gradient, then, with perhaps dithering on one extreme and large, flat planes on the other. Similar to how banding has its own spectrum of very obvious to less obvious. So then, how you find balance between the two spectrums, I think that would be a useful discussion.



Yes that's a very clever script, thanks for making it. I will check out Grafx2, especially since it has a linux build. Oh, also

Quote
With sharper screens, the pixels in AA are more apparent, but with higher-pixel-density screen, each individual pixel is less apparent.

This I follow, but I guess I had the underlying assumption that people aren't viewing pixel art at 1x on our modern screens. I always thought that pixel art was designed for smaller resolutions, even today where the games that use it either scale it up or have a low native resolution, which would make AA more important, since screens are indeed sharper.

Offline Ai

  • 0100
  • ***
  • Posts: 1057
  • Karma: +2/-0
  • finti
    • http://pixeljoint.com/pixels/profile.asp?id=1996
    • finticemo
    • View Profile

Re: NEW CLUSTER STUDY THREAD!

Reply #146 on: January 20, 2015, 01:40:18 am
Quote
In simple terms, regularization is tuning or selecting the preferred level of model complexity so your models are better at predicting (generalizing). If you don't do this your models may be too complex and overfit or too simple and underfit, either way giving poor predictions.

So if I'm getting this right, this is a less complex model, because it doesn't account for information that single clusters would. Or is it more complex because it has more rules?
Hmm. It helps to understand spline math, or gaussian blurring, I guess. There is the concept of quality of fit, but more important in this case, there is the idea of how well fitted things are to the things surrounding them.
For example, a conic spline is fitted to 3 variables, a cubic spline is fitted to 4.. In practical terms, fitting features to other features tends to mean that the features 'interlock' tightly -- many points on each feature align (along a line or more commonly a curve) with many points on other features, producing an image that feels very integrated and solid. The anisotropic smoothing filter I brought up tends to move an image towards this ideal (with the obvious downside that you may not want -all- features equally regularized.)

So I would say that this is a model that fits more variables, but also intentionally discards many variables (eg. those that you might find in trying for precise fidelity to reference), to create a whole that is much more obviously unified and beautiful (in a literal sense, but also perhaps in a mathematical sense)

Quote
Yes you're right, fidelity is selective. I guess coming from a traditional background I'm so used to it being the point, but I'm starting to see how it's not the point with pixels.

Quote
What you're saying is defragmenting is even more the point than banding elimination, that banding is almost accessory. I think I can see that.
Yeah, I like to kill banding, but realistically, it adds maybe 2-5% rendering quality, and the conclusion I think we've more or less agreed on in this thread is that solid clusters constitute 70-80% rendering quality -- which is enough to make a picture hang together, and coming back and doing debanding and AA is something that can go into a separate pass without noticable influence on how the assets work.

Quote
Yes, if you were to take it a step further it would then be a matter of reducing the clusters themselves (like with the lips). Interesting... this makes a sort of gradient, then, with perhaps dithering on one extreme and large, flat planes on the other. Similar to how banding has its own spectrum of very obvious to less obvious. So then, how you find balance between the two spectrums, I think that would be a useful discussion.


I'm not sure what to contribute to that discussion, but I agree.

Quote
Quote
With sharper screens, the pixels in AA are more apparent, but with higher-pixel-density screen, each individual pixel is less apparent.

This I follow, but I guess I had the underlying assumption that people aren't viewing pixel art at 1x on our modern screens. I always thought that pixel art was designed for smaller resolutions, even today where the games that use it either scale it up or have a low native resolution, which would make AA more important, since screens are indeed sharper.
Hmm. I may be a bit out of touch here. I am aware that some games scale up their art, but I had no idea how prevalent this was.
I would suggest that the effect on scaled up pixel art is slightly different, more like a simulation of gaussian blur, from my own subjective experience; it imparts a much more ambiguous type of smoothness the greater the effective pixel size. I find it adds something, but a different something than AA on lower-res / blurrier displays.

It's good to revive this discussion, hopefully others will get involved again soon.
If you insist on being pessimistic about your own abilities, consider also being pessimistic about the accuracy of that pessimistic judgement.

Offline Helm

  • Moderator
  • 0110
  • *
  • Posts: 5159
  • Karma: +0/-0
    • View Profile
    • Asides-Bsides

Re: NEW CLUSTER STUDY THREAD!

Reply #147 on: January 20, 2015, 09:07:07 pm
First of all, lovely art you've got there, Joe. I can look at it and immediately see how the process has helped you.

As to answering your questions,

Quote
Wwhich, in your mind, takes precedence between the three? Banding, Defragmentation, or Fidelity? I have a feeling you'd say it's on a case by case basis. My personal tendency is to prioritize banding-elimination.

Removal of banding gets a bit metaphysical at some point, where just by the process of stripping it out I often feel I am arriving at something more robust, but robust in what sense I can't really tell you, it's not making the subject clearer only, or adding detail (often it is removing detail). The way I would liken it is that zen quality of getting something down in painting with few colors and few brush strokes, it's not even exactly 'control' what I am describing here, more like chiselling something until it arrives at some quintessential level. Of course I can't explain this much better, suffice to say I've been working professionally, using most of my techniques (more on this later) to the point where they're second nature and they never interfere with getting things done to the best of my capacity, they only give an added self-sustaining enjoyment of getting things down to that quintessential level.

Defragmentation as you describe it I sometimes fight against consciously, adding more shades and blurs and dithers when the subject matter seems to desire it although that creates more softness and complication where a clear cut cluster could best suit. It isn't very often that this happens, though.

Fidelity, honestly for small-ish arcade game sprites, single pixels will happen, but aside from that I've never felt that I am sacrificing fidelity by banding removal and cluster clear up. The more you practice, the more it'll become second nature, there are wonderful solutions for most problems, and then there's that 'quintessence kick' when you find them.

Quote
2. Why does this work? What is it about the absence of orphan clusters that makes good pixel art?

I think it's flow, two (or more) married pixels always have a direction, as a brush stroke. An orphan pixel is a directionless dot, it has its uses but it doesn't suggest motion or connection between parts. When the contrast of the piece is such that you can both see the connections of clusters but also see them as related by lightness and shade, then you have a very harmonious, moving, sculpted entity.

Quote
3. AA was discussed earlier in the thread, as something that is less necessary. Wouldn't sharper screens make AA more necessary than before? Or did I misunderstand that point

No because AA relies on blurry crts to buffer a color in a color. The end result is that the blurry CRT art looks like painted art, convincingly, not that it looks like good, crisp pixel art. When you look at pixel art on a sharp screen, you're never going to hide the fact that the art is made of pixels as effectively as on a bleeding CRT monitor, so why try? Show pixel lines (esp, perfect lines), show connections, keep it crisp.  Would a game of tetris look more beautiful if you ran it through a gaussian blur filter? Digital art gets to be super crisp, this is an asset, why treat it as a deficit?

Quote
4. After a year of practice, where do the single and double pixel stand in your mind? As I understand it this exercise is meant to strengthen the integrity of your clusters, so that the single pixel becomes a tool, not a crutch. But there were many places where I knew single pixel AA would be the best solution, I merely refrained. I'm sure you don't think single pixels should be abolished; where do you feel they're appropriate?

I don't adhere to a no-single pixel rule in my daily professional artwork. It's just not practical for very 1. low-res artwork that is also 2. very cartoony stuff which is what I'm employed to do at the moment. I do try to minimize single pixel usage unless it's highlights, dot eyes, that sort of thing. Aside from that, my point of view is largely unchanged from when I last wrote on this thread: banding removal/defragmentation of shapes is the bedrock of my  technique and I think it's the only honest piece of craftsmanship that should be communicated and practiced for the learning pixel artist. The rest is just fluff that we created (for good or worse) in a community bent on classification and enumeration of 'techniques'. This is very common, most creative online communities create jargon. I do not return to 'selout', for example, as an effective and intersectional technique but I do think about banding and pixel shapes every day, in my work.

Tangentially: my recent interest is in bit restrictions for palettes, esp. 3 bits per channel, like the Atari ST. I use this restriction in my daily work because it explains the color choices of past videogame artwork very effectively, a lot of linear from-hue-to-different hue ramps, full saturation, less use of gray, less use of interchangeable colors, a stronger foundation in pure black for segmentation of forms and so on. This is a blind spot in how we look at art in Pixelation, because we jumped into learning pixels with a 24bit palette in our hands and no matter how restricted and small the palette we used was, it was still picked from a color space of millions of colors. So many close shades, so many earthly greys, so on. It is a joy for me to un-learn, for a time at least, my dependence on shifting colors through grey buffers to get things done, and it's really no way to practice this with the c64 palette or the EGA palette, honestly. I encourage anyone with the means to restrict their bits-per-channel to give this a good go, especially with 3,3,3 and 4,4,4.

« Last Edit: January 20, 2015, 09:10:36 pm by Helm »

Offline ptoing

  • 0101
  • ****
  • Posts: 3063
  • Karma: +0/-0
  • variegated quadrangle arranger
    • the_ptoing
    • http://pixeljoint.com/p/2191.htm
    • View Profile
    • Perpetually inactive website

Re: NEW CLUSTER STUDY THREAD!

Reply #148 on: January 21, 2015, 09:14:46 pm
The 333 and 444 thing is very true, fully agree there. Moreso 333. 444 is a good step closer to super smooth, esp if you have as many colours in your picture as you want. 555 is where you have more colours than you could ever want for pixelart in most cases.

I personally also have doen a lot of RGB222 (full 64 colour EGA palette, Sega Master System for example) stuff which is good fun.
It forces you to make bold choices with your colour usage.

About AA making less of a point on modern displays I only agree partly. It depends at what magnification you are looking at the work.
If you are making an application with pixeled icons that are looked at at x1 then yes, AA helps for sure. And depending on how high the PPI on your display is even at x2 and x3 the aa will hold up and soften things (esp curves), but I agree that it is pointless to try and make everything super smooth, or what I would say turns into overAAing things, which is something you can for sure get away more on a blurry CRT. But at the same time on a CRT you DO need less AA to get to the same level of smoothness than on a LCD.
« Last Edit: January 21, 2015, 09:34:08 pm by ptoing »
There are no ugly colours, only ugly combinations of colours.

Offline Joe

  • 0010
  • *
  • Posts: 298
  • Karma: +2/-0
    • View Profile

Re: NEW CLUSTER STUDY THREAD!

Reply #149 on: January 22, 2015, 12:39:29 am
@Ai: That makes more sense to me, thanks.

@Helm: Thanks! And yes, it really clicked when I tried it.

Stroke direction, that's probably it. I also see that single pixels are noise in most places. It definitely feels brushy when the clusters are aligned.

Obviously this isn't the only way to pixel, much like paint has many ways of being applied. I see some artists like tomic pull off dither very convincingly even on my crisp screen, and I would like to add that to my toolkit. But I am convinced this is the best way to pixel at present, barring any stylistic intentions.

The use of AA—like dither—is subjective, I don't see any conclusive evidence that it has less of a place. I say this with the assumption that over-AA was never appropriate. Here is an arbitrary curve:



Clearly, the buffer color is there; it's not fooling anyone. But it's a trade-off, and I think the benefit outweighs the cost, that the curve is effectively smoothed. Yes you could convert the curve to perfect lines, but that's not always an appropriate solution. That said, thanks for clearing up why LCDs reduce the effectiveness of AA.

On bit restrictions: That's been mentioned here before, but I haven't seen it discussed or described in depth. If you feel strongly about its usefulness for learning, why not start a ramblethread in that vein? I want to try but don't know where to start. Also why would bit restriction exclude grays? Or is that just personal choice

@ptoing: Are there any introductory articles you'd recommend on the subject? Don't know how you would calculate whether a color was in that space, how you would set up the palette, etc. Feel like something that fundamental should be in the linkage thread.

I revisited the Altered Beast sprite with interesting results. Less cautious and more daring is what comes to mind of how I felt while working. There were many solutions I never would've considered before. Original -> Old -> New: