Anybody have an RTX video card? Nvidia's "Canvas" has been released to the public!

[OysterMug]

Anybody out there drop a buttload of money on an nVidia RTX video card and would like to see what it can do for TK17? ?

 

nVidia has finally released their "AI enhanced" image software called "Canvas" to the public that can turn a simple doodle into a photo-realistic image.

 

I'd LOVE to see what it could do with an already "lifelike" VX screenshot.

 

[dxw]

Very fancy healing brush! I would probably use this over the online offerings for generating fuzzy shapes --- but only if I wasn't aware of faster ways to do so in PS already.

 

Such tools are functionally useless without the ability to tag your own reference images. There is no "AI" here, only machine learning. The generative effect doesn't come from algorithms that actually define materials. The Canvas tool has no concept of opacity, depth, etc. You can paint on multiple layers to almost no effect.

 

You can see the limits of GAN tools by way of how quickly they lighten or darken an image as you add detail. I found that using more than ~5 brushes would cause additional strokes to start canceling each other out, creating pitch black areas on the image where "gravel" was supposed to be, etc. Unless the word "material" starts to mean something more than "best guess at pixel colour", we'll hit another road block with these tools even if they begin to allow users to tag their own images.

 

You can achieve more controlled results, quicker, in PS. I'm at best an amateur, but this technique has allowed me to "paint landscapes" in minutes in whatever resolution I want:

 

1. Start by laying down colours as the Canvas tool encourages you to.

2. Find a reference image (rocks, sky). Anchor the healing brush to the point of interest.

3. Change your brush's blending mode to a value-sensitive one (luminosity; color; etc).

4. Heal-brush over your colours. Watch the texture transfer.

5. Use Content Aware fill to close seams, gaps, and areas with detail you aren't sure about.

 

That sounds like a long process because it's written down, but in reality, it's faster than starting with Canvas if you have even a couple reference images laying around. The argument that Canvas' output is supposed to cleaned up in Photoshop defeats its value proposition. Why not just start and end in Photoshop with more control? I've already got contrast-sensitive brushes sampled from nature, so why would I use a generative texture stamping tool if my brushes already... paint... the same...?

 

I don't want to take away from the neatness factor -- painting with ML brushes is a great idea! But recent versions of Photoshop are shockingly effective at smart fills and already do this sort of generative thing with much better results. Most people aren't aware of this because they're either still using a pirated copy of CS6, or the sheer size & complexity of PS turns them off from learning.

[KoolHndLuke]

1 hour ago, dxw said:

size & complexity of PS turns them off from learning

True. I've been dabbling in PS for years and stumble across stuff I didn't know all the time, then work on something else and forget all about it. Only so much info I can keep in my old brain and I think I have a memory leak, lol. Pretty sure my approach is off because I thought I could pan out and take it all in with regards to modding. Unfortunately there's just too much for one person to know and I find that I really need to concentrate in more specific areas like learning more about PS.

[dxw]

On 6/26/2021 at 8:43 PM, KoolHndLuke said:

I think I have a memory leak, lol

 

It's not you -- it's "the times". I've been watching a trend -- as the value of digital art goes up, the complexity of the tools increases disproportionate to the level of benefit that updated tools provide. Job security code > streamlined features.

 

For example, this Canvas app -- nVidia could have given users the ability to train/tag their own source material. There's no reason for them to omit such a feature in beta software, other than to manipulate user expectations. Teasing gets old quick in software .

 

Or, take Photoshop: the default brushes are optimized to drive users insane. You switch your brush type, everything changes. This isn't necessary. To fix the problem I deleted every single brush in CC and started over. Attached is the core set I use. Now I can select a basic shape and size and switch tools without having a seizure. When I discovered I had to make my own triangle brush because Photoshop doesn't include a no-frills one, I facepalmed, and nearly refunded the software.

 

I wish more development time was spent on making the existing tools we have work really well, instead of racing forward with machine learning. If Photoshop becomes so complex that it needs a rewrite, Canvas won't matter. Eh.

SomePSCCBrushes.abr

[judge007]

On 6/25/2021 at 5:49 AM, OysterMug said:

Anybody out there drop a buttload of money on an nVidia RTX video card and would like to see what it can do for TK17? ?

 

nVidia has finally released their "AI enhanced" image software called "Canvas" to the public that can turn a simple doodle into a photo-realistic image.

 

I'd LOVE to see what it could do with an already "lifelike" VX screenshot.

 

 

12 GB RTX 3060

Thanks!