Topaz AI Gigapixel

[uhohitsuhoh]

If you are an artist or developing a game this is going to be a must have. I'm not super well versed on the tech behind it all, but basically it can enlarge an image without losing any of the quality, and even improving it in most instances. I found out about this because I found an article about people remastering final fantasy 7, apparently the backgrounds in that game were pre-rendered images that were rendered very small, so using this program they were able to not only increase the size but bring much more quality to those images.

You must be registered to see the links

Not sure how to embed links but that vid is an overview of the program.

Now its generally known that you can increase the size of a render, like if you render in 4k, and then resize the render afterwards to 1080, it can clean it up a little, just because of whatever the post processing work that happens, again i'm illiterate in this stuff whether its nearest neighbor blah blah. but basically you can get a quality render quicker using this method.

With this program though, you might be able to do even quicker renders, by rendering in just 720p, upscaling it 600 percent, and then downscaling it to 1080, because you are gaining quality in the upscale as well as the downscale.

This means less rendering time, but it does mean more post processing time. However, like batch rendering you can batch image scale numerous photos quickly however, without tweaking settings for individual images etc.

Thought I'd share this with anyone to mess with to help shave rendering times

edit- also I failed to mention its a 100 dollar program, not a free one. I caved and purchased it because I couldn't find it online yet lol

 

[Passenger3D]

It upscales stuff quite quickly since it uses gpu processing but its not THAT great as they claim. It depends on the situation, its good at some thing at others not so much. Its kinda hit n mis. So its not some magical fix, if you want quality renders your best bet is still to actually render at high resolution.

It works decently for high res upscaling like fullhd to 4k or 4k to 8k but even then you are not getting any additional detail(that would be there if you rendered at 4k) its just that it doesnt seem you are losing the detail from upscaling so its kinda you get fullhd at 4k its still crisp but real 4k would be better.

It doesnt work too well for low res to high res, so if you are hoping to render at 50% of fullhd and then upscale to fullhd im afraid its not gonna be great simply because in order to upscale properly AI needs details to work with thats why high res upscaling(fullhd to 4k) works better then lowress to highres.

so in the end its not really adding detail it just kindof inteligently upscales it so that it seems as if you didnt lose detail

bottom line - yes its good, no its not as great as they claim it to be and no its not a magical fix. gl

 

[Winterfire]

If anything, a common quick fix to grainy effect should you not let the iray run as much as it should is to scale, so that software probably does a better job with that

 

[f95zoneuser463]

You must be registered to see the links

and I have good news:

As of right now the software works even after the 30 day trial should be over.
Not sure why that is, could be a bug, could be due to the updates. I know there a ways to "defeat" theses typical time limited trial version, but I haven't done anything like that.

Meanwhile the software got a few updates. The interface changed and bugs that I've noticed in older versions are fixed too.

The software also handles huge amount / long image sequences better (still needs a lot of work) ... this is important because people would love to use this for video-upscaling too. I have done this already, but you MUST know what you are doing. Picking the wrong settings for a image-sequence with 10000+ images can result in process-time of days instead of an hour!
One BIG problem when using it for video/long image sequences: is that the software processes everything in a single thread.
The speed would improve if the devs implement a second thread for writing the upscaled image to the disc while the GPU already processed the next image.

There are other projects that work with similar technology, but for real-time videos. This software got me thinking a lot about "remastering old videos". I searched the net for realtime AI upscale software, found some things, but nothing public yet.

What the software should NOT be used for is denoising! I tried that too, but it's 'wrong thinking'. It's not what the software is build for. There are better algorithms and tools specifically for that. In general it's EXTREMELY important to have a high quality source image. Compression artifacts in the source will only be amplified. Typical JPGs from the web with high lossy compression are a bad source.

If people are interested I can upload a video comparison for upscaled video with this software. It's the intro from 2010 "My Plaything Audrey Bitoni". It's a later DVD generation, digitally recorded and mastered, not the early 2000s stuff where they just sold VHS converted to DVDs, that's important for quality. I've upscaled it 200%.

Here is a

You must be registered to see the links

I found from someone else. Check his reply in the comment section, he used AI Gigapixel too and he clearly knows what he is doing.

Funny and highly recommend talk about image upscaling with neural networks:

You must be registered to see the links

Other projects:

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

(not upscaling)

You must be registered to see the links

While researching I also found a software called

You must be registered to see the links

. It was used to restore the 1927 silent movie "

You must be registered to see the links

". I believe this was after the missing scenes where found.

Exciting times, I can't wait for the day where you can play some old videos, press the "enhance" button and it looks so much better. That will happen, I'm sure.

 

[uhohitsuhoh]

Yea I didn't mean to give off the impression that this would fix a grainy image, I was just excitedly posting this quickly after messing with it last night.

I personally have a 6gb 980ti, and I'm working on my first VN. And by first I mean its my tenth attempt at a first lol. Anyways I'm working with lower light enviornments and I'm still super noobish with how it all works, Without using ghostlights and emissives that completely wash out the image and details I continually get minor noise in the renders. I'm also rendering in full 4k. So I've let problem renders go for 10 hours while sleeping, just a single image and STILL got white specs on figures and the like.

I use a scene optimizer for DAZ3d, as well, since I only have 6gb, so if I want multiple figures in a scene, I have to downscale everything possible other than the figures. The enviornment and textures etc. So I'll have a pristine looking figure in a blah enviornment alot of the time.

Using this though, I can take that 4k image and upscale it by 6x/600 percent, and it actually makes the enviornments look much closer to how they shouldve prior to downscaling, while also making the figures look much more crisp.

This is all after I've run a version of a denoiser however. I use photoshop, and since I'm rendering at 4k and not cancelling anything early the only noise i get are these strange white specs, which get taken care of by using "median". This seems to be the least harmful option and by my untrained eye its basically just deleting the white specs themselves, but if you zoom in enough you can see the telltale blurryness. It isn't much but its there.

That also gets taken care of in the upscale

Then, when its downscaled back to 4k or 1080, I get the sharpness benefits from that as well.

I've deleted my original renders and am only left with my final images, but I'll try and get up a few from each stage of the process to show what its doing for me personally. I probably won't post a 6x 4k image however because its something ridiculous like a 280 mb png file, lol, but i'll post a normal 1080 render untouched, a 1080 render after the median denoising, and a 1080 render after median denoising, upscaling and downscaled back to 1080.

 

[f95zoneuser463]

I use a scene optimizer for DAZ3d, as well, since I only have 6gb

Oh I remember painfully optimizing every little detail because I started on a 760 GTX 2 GB. Scene optimizer was required for almost every scene. But hey, looking back at it now: I learned how optimized the scenes and many tricks to free up as much VRAM as possible. Even my freaking Ren'Py launcher is configured to run in software rendering to not consume VRAM haha.

I'm also rendering in full 4k. So I've let problem renders go for 10 hours while sleeping, just a single image and STILL got white specs on figures and the like.

We are both doing the same thing then. 4K renders at night. Have you tried the DAZ 4.11 beta with AI denoiser to fight the remaining 'white specs'? I don't know how 'usable' it is on a 6 GB card since this will only increase the memory hunger. I'm on a 1070 GTX 8 GB now and even that's not enough in some cases.

concerning AI Gigapixel:
The quality is obviously questionable and I think it should not be used where it's not necessary. I'm not even that interested upscaling actual 3DCG. Personally I'm not limited by rendering times. The time required to script the VN and create the 3D scenes in the first place is long enough to render in the background. It's probably different for every dev and depends on the workflow.
The thing that excites me: having the option to "enhance" old images and especially videos in cases where simply no better source exists. DVDs are an excellent example. They have a "good enough" resolution. Assuming you could run this in realtime on a movie ... I bet you could trick people to think that it's a 1080p video ... not at a blu ray quality level, but recorded on a phone 1080p cam.

 

[depechedNode]

Necroing this useful thread is necessary, this software deserves attention.

I'm using Gigapixel AI for a while and I must say I'm impressed after using Waifu2x for a long time. It's not a miracle software but very close, as mentioned above posts the higher the resolution better the upscaling gets. It needs more improvement, more optimisation, but it's getting there, currently its the best I used. I'm not a dev, I use it for my photos, upscaling and/or denoising some games.

Mostly I use it for denoising (I hate noise in renders), it depends on the personal taste but I'm very happy with the results, it's much better than Waifu2x. I know waifu2x's main focus is drawn art, illustrations, etc, but it still did a good job with 3DCG art too. Gigapixel also does a very good job with drawn art too, but it's slower than waifu2x.

Upscaling only works good if the image is clear, crisp, sharp enough even if it's low resolution. If the image is downscaled wrong, very aliased, very low bitrate, filled with artifacts, extremely blurry, etc than the result is bad usually.

One of the big current downsides of Gigapixel is you can't use it as a background tool (even if there's a setting). You may use it but you must not do anything requires GPU usage, even browsing web can be painful due to Gigapixel sucking too much GPU power. Waifu2x handles this much better, it also sucks all GPU power but not 100% like gigapixel, it gives headroom for other softwares to be able to use GPU. This needs to be improved by letting Gigapixel to use between 95-98% max GPU usage or having a lower priority.

This not a Vram issue as some may think, because used GPU is a 1080ti and low or high Vram setting makes no difference. You can even watch a movie while using Waifu2x, but with Gigapixel even working on a word document can be painful. Only alternative currently is using CPU, but as you may guess it's much slower than a dedicated GPU even with a 9xxx series CPU.

Some tips for more speed:

- Disable "Face Refinement", useful only for very low res group photos.
- Disable "maximum quality AI models" if you're working with 3DCG, tho it's useful with real photos, there's a quality difference even it sometimes creates some artifacts.

Tips for denoise and upscaling, this actually depends on the personal taste, so test some settings and find your own settings or be lazy and just enable auto setting.

- For upscaling real photos I prefer 0.25-0.50 denoise and 0.40-0.70 remove blur
- For denoising and/or supscaling 3DCG 1.0 denoise and 0.50-0.80 remove blur

For using Gigapixel AI only for denoise, choose custom upscaling and set the value to "1".


And here are some examples (AI models off):

- Original vs 4x upscaling:

You must be registered to see the links

You must be registered to see the links

- Only denoise:

 

[MrHotCoffee]

Necroing this useful thread is necessary, this software deserves attention.

I'm using Gigapixel AI for a while and I must say I'm impressed after using Waifu2x for a long time. It's not a miracle software but very close, as mentioned above posts the higher the resolution better the upscaling gets. It needs more improvement, more optimisation, but it's getting there, currently its the best I used. I'm not a dev, I use it for my photos, upscaling and/or denoising some games.

Mostly I use it for denoising (I hate noise in renders), it depends on the personal taste but I'm very happy with the results, it's much better than Waifu2x. I know waifu2x's main focus is drawn art, illustrations, etc, but it still did a good job with 3DCG art too. Gigapixel also does a very good job with drawn art too, but it's slower than waifu2x.

Upscaling only works good if the image is clear, crisp, sharp enough even if it's low resolution. If the image is downscaled wrong, very aliased, very low bitrate, filled with artifacts, extremely blurry, etc than the result is bad usually.

One of the big current downsides of Gigapixel is you can't use it as a background tool (even if there's a setting). You may use it but you must not do anything requires GPU usage, even browsing web can be painful due to Gigapixel sucking too much GPU power. Waifu2x handles this much better, it also sucks all GPU power but not 100% like gigapixel, it gives headroom for other softwares to be able to use GPU. This needs to be improved by letting Gigapixel to use between 95-98% max GPU usage or having a lower priority.

This not a Vram issue as some may think, because used GPU is a 1080ti and low or high Vram setting makes no difference. You can even watch a movie while using Waifu2x, but with Gigapixel even working on a word document can be painful. Only alternative currently is using CPU, but as you may guess it's much slower than a dedicated GPU even with a 9xxx series CPU.

Some tips for more speed:

- Disable "Face Refinement", useful only for very low res group photos.
- Disable "maximum quality AI models" if you're working with 3DCG, tho it's useful with real photos, there's a quality difference even it sometimes creates some artifacts.

Tips for denoise and upscaling, this actually depends on the personal taste, so test some settings and find your own settings or be lazy and just enable auto setting.

- For upscaling real photos I prefer 0.25-0.50 denoise and 0.40-0.70 remove blur
- For denoising and/or supscaling 3DCG 1.0 denoise and 0.50-0.80 remove blur

For using Gigapixel AI only for denoise, choose custom upscaling and set the value to "1".


And here are some examples (AI models off):

- Original vs 4x upscaling:

You must be registered to see the links

You must be registered to see the links

View attachment 512937 View attachment 512938

View attachment 512959 View attachment 512960


- Only denoise:

View attachment 513008 View attachment 513009

View attachment 513020 View attachment 513021

Hey, so I have started working with Daz a week-ish ago and now I'm starting to learn lighting correctly, I have a question.

My idea of rendering is: What if I render an image in 4k, use gigapixel AI to even improve it further and then downscale it to a size usable in ren'py. Is that a good idea? If yes, how should I downscale it? Which software?

Thank you in advance

 

[depechedNode]

Hey, so I have started working with Daz a week-ish ago and now I'm starting to learn lighting correctly, I have a question.

My idea of rendering is: What if I render an image in 4k, use gigapixel AI to even improve it further and then downscale it to a size usable in ren'py. Is that a good idea? If yes, how should I downscale it? Which software?

Thank you in advance

It totally depends on the quality of your render, if the renders are gonna be a little noisy, blurry then it can be a worthy effort to denoise and upscale, otherwise it maybe just an overwork for a good 4K render. So choice is yours if you see an improvement with the result.

This AI processing sometimes brings interesting results for 3DCG visuals. For example while making the overall visual sharper, human skin looks softer and I like this softer look more.

Downscaling is not a big issue, you can use a tool like XNConvert (batch processing support) or any other tool you want.