[Jimmy]

Yeah I'm talking about actually fixing errors in the original art. The gun barrels in most of those tiles show erroneous AI behaviour because the original art was so lazy.

[Zaxtor]

Nice stuff.
Nothing wrong with improving stuff.
Those mobs are from mid 90s so making 2010s stuff is good.
Improvement etc

lots of old stuff get improvement.

[Phredreeke]

Jimmy 100MPH, on 19 February 2019 - 10:38 PM, said:

Yeah I'm talking about actually fixing errors in the original art. The gun barrels in most of those tiles show erroneous AI behaviour because the original art was so lazy.

TBH I don't think it matters that much if those edits are made pre or post upscale.

Here's another pair of frames, this time showing one of the flaws in the process. some of the smaller blood spatter was erased while creating the mask. (frames like these I would edit by hand in my Blood upscale)

[Danukem]

Still damned good, even if missing a few pixels of blood. I would argue that the Octabrain upscale is good as-is: the tendrils of blood thin out in a very natural way and look fine, even if it's not 100% faithful. If a human being had produced that image a few years ago, I doubt that anyone would have complained.

The enforcer on the other hand needs a little help where the arm is coming off.

[Jimmy]

Fixing bad behaviour after the upscale sounds like 10x the work to me.

[Phredreeke]

I reworked the way I made masks, using a pair of upscales as a basis, instead of the original sprite.



Left new, right old. It looks better for these sprites, but worse for others. Possibly I can filter out part of the image and overlay it on the old mask...

[oasiz]

These are by far the most promising Duke3D upscales I've seen.

[Radar]

I don't know anything about upscaling. Is this something that can be done on-the-fly at decent frame rates, or is an asset drop-and-replace the only way this can be implemented?

[MusicallyInspired]

No. At least not yet. It takes anywhere from one to several seconds to generate a single image depending on how large the source image is (it increases by 4x) and how fast/how much memory your GPU has. It can be done by CPU as well but the processing time increases quite a bit.

[Phredreeke]

Honestly I don't think you'll ever get low enough latency to do it in real time.

[MusicallyInspired]

Yeah and we've never needed more than 640k of memory, right? You underestimate technology. We've got real-time raytracing now that looks ten times better than Toy Story's pre-rendered graphics from 1995. The Starry Expanse remake project looks better than Riven's 1997 state of the art pre-rendered graphics. It'll happen. I give it 10 years max. And that's generous.

[Phredreeke]

You're right, I'm being overly cynical, also I'm not considering application specific implementations, like cacheing previous frames.

[Jimmy]

Consider this: One day you'll be able to feed sprite frames into an AI that constructs a completely accurate 3D model out of it. And it won't be long now either.

[MusicallyInspired]

You know what would be neat, create a sprite from 3 angles (front, back, side) in the standard t-pose. Then the algorithm would extrapolate it into a 3d model and animate a standard walking cycle and spit out the appropriate animation frames in whatever resolution you want.

[Radar]

I just want polymer lights without dropping down to 10 fps, k.

[MusicallyInspired]

Get a faster system. Mine works fine. I go below 30 sometimes, but I've never had major problems otherwise with the HRP in the standard Duke levels. Ever. I don't ever go down to 10.

[Radar]

I'm kidding around. I actually don't know how fast polymer runs on my system. (probably slow though)

[Mark]

Jimmy 100MPH, on 20 February 2019 - 02:41 PM, said:

Consider this: One day you'll be able to feed sprite frames into an AI that constructs a completely accurate 3D model out of it. And it won't be long now either.

That goal is close already. At least a year ago I saw vids of programs that turn a series of still pics into a 3d model. It required more than 3 views but they work very well. I remember downloading one of the programs but never tried it. I must have gotten sidetracked with something else and forgot about it until now.

Some modeling programs already have an "autorig" function to rig up an animated skeleton to a model. They are not perfect and rely on some tweaking afterwards. But the main parts of the dream program are here now just waiting to be refined and combined.

[----]

Mark, on 21 February 2019 - 05:53 AM, said:

I must have gotten sidetracked with something else and forgot about it until now.

\o/ Boobies

[Phredreeke]

fuegerstef, on 22 February 2019 - 11:27 AM, said:

\o/ Boobies

[Micky C]

One difficulty in turning the Duke sprites into models with ANNs is that the sprites are quite noisy in that they’re touched-up versions of images of the original models. The ANNs might have some difficulty in overcoming discrepancies between the different views. I’m sure it’s possible, but it makes the task more complicated.

Add to that we probably want high quality models, not just low-res 3D versions of the original art. I’d guess upscaling 3D models is a lot more difficult than upscaling 2D images?