MHLoppy

MHLoppy, 1 year ago

"Open source AI" is apparently the answer.

MHLoppy, 1 year ago (edited 1 year ago)

Yes, though just nitro basic. Discord doesn't show ads and claims to not sell my data. While I can afford to do so, I'd much rather pay a few bucks a month to keep it that way.

The number of people in this thread aggressively against a free-to-use service having any kind of way to pay employees and server bills makes me fucking depressed, and helps to explain why most free services I enjoy never seem to stay afloat with just an optional payment-based membership thing.

Edit: To people suggesting less corporate-based (whether FOSS or not) alternatives, that's totally cool! Just remember that the people behind these projects need some way to pay the bills the same way the corporate ones do, so I encourage you to contribute to them, whether that's through e.g., code improvements (which doesn't pay bills but is still helpful!) or plain old donations.

MHLoppy, 1 year ago

It depends a lot on what's being encoded, which is also why different people (who've actually tested it with some sample images) give slightly different answers. On "average" photos, there's broadly agreement that WebP and MozJpeg are close. Some will say WebP is a little better, some will say they're even, some will say MozJpeg is still a little better. Seems to mostly come down to the samples tested, what metric is used for performance, etc.

I (re)compress a lot of digital art, and WebP does really well most of the time there. Its compression artifacts are (subjectively) less perceptible at the level of quality I compress at (fairly high quality settings), and it can typically achieve slightly-moderately better compression than MozJpeg in doing so as well. Based on my results, it seems to come down to being able to optimize for low-complexity areas of the image much more efficiently, such as a flatly/ evenly shaded area (which doesn't happen in a photo).

One thing WebP really struggles with by comparison is the opposite: grainy or noisy images, which I believe is a big factor in why different sets of images seems to produce different results favoring either WebP or JPEG. Take this (PNG) digital artwork as an extreme example: https://www.pixiv.net/en/artworks/111638638

• Original: 6,080,273 bytes
• MozJpeg (Q88, 4:2:0): 438,687 bytes
• WebP (Picture preset, Q90): 595,634 bytes

This image has had a lot of grain added to it, and so both encoders end up with a much higher file size than typical for digital artwork at this resolution. But if I put a light denoiser on there to reduce the grain, look at how the two encoders scale:

• MozJpeg (light denoise, Q88, 4:2:0): 394,491 bytes (~10% reduction)
• WebP (light denoise, Picture preset, Q90): 424,612 bytes (~29% reduction)

Subjectively I have a preference for the visual tradeoffs on the WebP version of this image. I think the minor loss of details (e.g., in her eyes) is less noticeable than the JPEG version's worse preservation of the grain and more obvious "JPEG compression" artifacts around the edges of things (e.g., the strand of hair on her cheek).

And you might say "fair enough it's the bigger image", but now let's take more typical digital art that hasn't been doused in artificial grain (and was uploaded as a PNG): https://www.pixiv.net/en/artworks/112049434

• Original: 2,184,811 bytes
• MozJpeg (Q88, 4:2:0): 200,933 bytes
• WebP (Picture preset, Q90): 164,078 bytes

Subjectively I once again prefer the tradeoffs made by WebP. Its most obvious downside in this sample is on the small red-tinted particles coming off of the sparkler being less defined, [see second edit notes] probably the slightly blockier background gradient, but I find this to be less problematic than e.g., the fuzz around all of the shooting star trails.. and all of the aforementioned particles.

Across dozens of digital art samples I tested on, this paradigm of "WebP outperforms for non-grainy images, but does comparable or worse for grainy images" has held up. So yeah, depends on what you're trying to compress! I imagine grain/noise and image complexity would scale in a similar way for photos, hence some of (much of?) the variance in people's results when comparing the two formats with photos.

---

Edit: just to showcase the other end of the spectrum, namely no-grain, low complexity images, here's a good example that isn't so undetailed that it might feel contrived (the lines are still using textured [digital] brushes): https://www.pixiv.net/en/artworks/112404351

• Original: 1,804,869 bytes
• MozJpeg (Q88, 4:2:0): 339,182 bytes
• WebP (Picture preset, Q90): 244,224 bytes

I quite strongly prefer the WebP version here, even though the JPEG is 39% larger!

Edit2: I've corrected the example with the sparkler - I wrote the crossed out section from memory from when I did this comparison for my own purposes, but when I was doing that I was also testing MozJpeg without chroma subsampling (4:4:4 - better color detail). With chroma subsampling set to 4:2:0, improved definition of the sparkler particles doesn't really apply anymore and is certainly no longer the "most obvious" difference to the WebP image!