Hello. I have a 96 x 96 bitmap and want to make it four times bigger. After it is this enlarged size I want to make all of it's edges smooth (probably using antialiasing.) I have Gimp and Inkscape and can't get any real sort of workings.
Is it antialiasing and how is it be done? The bitmap image is on a single color background, brown.
Thinking about it, I suppose it will be drawn on a single color background too!
Thanks,
Josheir
The operation you want is called "magnification filtering" or "resampling", not antialiasing. You can't avoid blockiness easily because the original image simply doesn't have enough data to uniquely fill the larger area, but you can choose a smooth filter like bicubic, to trade accuracy for visual smoothness.
In Gimp when you scale there should be a : Select Interpolation method. Try out a few and see if your happy, if not try downloading a new Interpolation method.
There is free software that focuses on this, but even they are limited.
If you want pixel perfect results, you should use no Interpolation method. That should just increase the pixel size. 1pixel -> 4pixels of the exact same color.
This allows screens to then scale it down again without breaking the image. If your making a pixel art game this is how you would do it.
There are several algorithms for upscaling pixel art: https://en.wikipedia.org/wiki/Pixel-art_scaling_algorithms
Maybe your image editing program of choice does already have some or you can download plugins.
On 2/2/2018 at 7:06 AM, Kryzon said:The cutting edge tech in image resizing is using neural networks:
Had to try it.
It just uses cubic and some noise that I think it collects from images. Really bad results on both photos and hand drawn art. You will get better results with Photoshop and some filters.
Maybe one day it will be good, worth watching.
Without you posting any proof I'm gonna have to call your bluff.
Someone else compared it with standard cubic resampling (using Photoshop, but cubic resampling is present in almost any image editor one can get). They had a different experience than yours.
"Photoshop 16x bicubic:"
"letsenhance.io 16x (ran twice):"
17 hours ago, Kryzon said:Without you posting any proof I'm gonna have to call your bluff.
With my first try I also used a image from the net, when I noted how near the two was I grew suspicious. The next two tests I ran using a image I drew and a photo of me.
I have no more images scaling left and I am not going to upload my photo, so here is only my hand made try with the "Boring" scale:
Original:
This is a jpg icon I used for a game I planned. It was very badly made as it was for planing, I can already feel the judgment.
Let's Enhance .io:
Now my result using Photoshop:
I used (Scale*2(Cubic)-> Remove noise)*4
Even without downloading you will see it's better. That wonky eye is really disturbing.
The black is the difference between my Cubic scaling and there scaling.
I am not uploading the photo, it was taken at work and we are not allowed to take photos of the studio also I am ugly. The results using the "Magic" scaling with my photo was even worse than the "Boring" scale. (Once I can use the page again I will take a photo and show it here as well.)
Conclusion:
My scale trick is lower quality than most free scaling software out on the net, you can find tools that scale much better than me. Even so I was able to do better. Let'sEnhance.io < My Cubic scale < Free scaling software.
Why we should watch this software:
The image I used from the google search was near perfect. Now this could mean they where cheating and used google search to find the larger image.
What I am hopping is that because there was more info for that image the results was better, because this would mean that with time the software will get better as the developers improve and the AI collects data.
For now it's good, but anyone with Photoshop could do better and most other free scaling tools are much better.
Thank you for sharing. Your Photoshop scaling method is interesting.
When you compare both versions (letsenhance vs ps), they have some slight shape differences (at the outer edges of the ears, for example). One of these versions might be more faithful to your original design (the one named Original), in case you downscaled it to make that thumbnail, so this is a way to tell which method is best.
If you drew it small like that then I don't think there's a reliable way to know which method is best, it's subjective.
I'm not sure how they trained their AI. I would guess they took HD photos, downscaled them and then told the AI "to take this low-res image input and generate this HD image output". You do that a lot of times until you have a generic model for that AI to upscale small digital photos.
Maybe their model works best with photography (pictures taken with digital cameras), rather than man-made images.
For example, someone trained an AI model to specifically upscale anime-style images: http://waifu2x.udp.jp/
28 minutes ago, Kryzon said:If you drew it small like that then I don't think there's a reliable way to know which method is best, it's subjective.
The 128*128 is the large image, it was made for a Icon of 64*64. It was a Tiny study on shading. Never knew it looked that bad so large.
39 minutes ago, Kryzon said:Maybe their model works best with photography (pictures taken with digital cameras),
As I mentioned I did try a selfie and the results wasn't good, blurry an noisy.
If the amount of images hasn't re-filled by this weekend I will donate to them. I want to see how good it is with outdoor images.
When you think about it there is more outdoor images on the net than indoor, so maybe it has better training with them. Either way I want to see what it really can do; this time I will be compare it to free software.
48 minutes ago, Kryzon said:someone trained an AI model to specifically upscale anime-style images: http://waifu2x.udp.jp/
Thanks for this one. A quick try with a rough sketch shows it is great. It produces good results very fast, so you only need to use it once or twice.
