Arbitrary-Scale Super-Resolution with Neural Heat Fields

I've been wondering exactly this for a while, if somebody more knowledgeable knows why we're not doing that I'd be happy to hear it.

Rather than operate on pixel space directly, they learn to operate on images that have been encoded by a VAE (latents). To generate an image with them, you run the reverse diffusion (actually flow in the case of flux) process they’ve learned and then decode the result using the VAE.

These VAE encoded latent images are 8x smaller in width/height and have 4 channels in the case of Stable Diffusion and 16 in the case of Flux.

I do think it would be more useful if it worked more like you said, though - if the channels weren’t encoded arbitrarily but some of them had pretty clear, useful human meaning like lightness, it would be another hook to control image generation.

To some extent, you can control the existing VAE channels, but it is pretty finicky.

I've never understood the DCT to be psychovisually optimal at all. At lower bitrates, it degrades into ringing and blockiness that don't match a "simplified perception" at all.

The frequency domain models our auditory space well, because our ears literally process frequencies. Bringing that over to the visual side has never been about "psychovisual modeling" but about existing mathematical techniques that happen to work well, despite their glaring "psychovisual" flaws.

On the other hand, yes a HSV color space could make more sense than RGB, for example. But I'm not sure it's going to provide a significant savings? I'd certainly be curious. It also might create problems though, because hue is undefined when saturation is zero, saturation is undefined when brightness is zero, etc. It's not smooth and continuous at the edges the way RGB is. And while something like CIELAB doesn't have that problem, you have the problem of keeping valid value combinations "in bounds".

Aren’t Thera’s frequency banks pretty darn close to DCT or Fourier transform already? This is a frequency space decomposition & reconstruction, and their goal is similar to JPG in that it aims to capture the low frequencies accurately, and skimp on the frequencies that matter less, either by being less visible or lead to error (aliasing artifacts). It doesn’t seem entirely accurate to frame this paper as learning in pixel space.

As far as perceptual color spaces, yeah that might be worth trying. It’s not clear exactly what the goal is or how it would help, but it might. Thera does use the same color spaces that JPG encoding uses: RGB and YCbCr, which are famously bad. Perceptual color spaces save some bits in the file format, and like frequency space, they are convenient and help with perceptual decisions, but it’s less common to see them used to save work, at least outside of research. Notably, image generation often needs to work in linear color space anyway, and convert to a perceptual color space at the end. For example, CG rendering is all done in linear space, even when using a perceptual color metric to guide adaptive sampling.

Another question worth asking is whether in general a neural network already learns the perceptual factors. When it comes to black box training, if the data and loss function capture what a viewer needs to see, then the network will likely learn what it needs and use it’s own notion of perceptual metrics in it’s latent space. In that case, it may not help to use inputs and output that are encoded in a perceptual space, and we might be making incorrect assumptions.

In this case with Thera, the paper’s goal may be difficult to pin down perceptually. Doesn’t the arbitrary in ‘arbitrary-scale super resolution’ toss viewing conditions and the notion of an ideal viewer out the window? If we don’t even want to know what the solid angle of a pixel is, we can’t know very much about how they’re perceived.

https://arxiv.org/abs/1907.11503

https://arxiv.org/abs/2308.09110

With generative ai they tend to have a learned compressed representation instead (VAE)

E.g. in the image you can see a diagonal bands basis function. Image codecs don't generally have those-- not because they wouldn't be useful but because codec developers favor separable transforms that have fast factorizations for significant performance improvements.

I don't think we know and can really make good comparisons between traditional tools and ML powered compression because of this. We just don't have decades of efforts where the engineers were allowed a million multiples and a thousand memory accesses per pixel.

I tried doing some pixelart->HD conversion with Gemini2.0Flash instead and the results look quite promising:

* https://imgur.com/a/t9F94F1

The images are however all over the place, as it doesn't seem to stick very close to the prompt. Trying to fine tune the image with further chatting often leads to overexposed looking pictures.

All the results are done with prompts along the lines of "here is a pixelart image convert it into a photo" or some variation there of. No img2img, LoRA or anything here, all plain Gemini chat.

It wouldn’t be more funny ha-ha, just more funny strange.

That said, there is a benefit to fine-tuning a model on a reduced data set after the initial training. The initial training with the larger dataset means that it doesn’t get entirely lost in the smaller dataset.

That said, our approach is actually trained on a (by modern standards) rather small dataset, consisting only of 800 images. :)

But it's extremely time-consuming and currently expensive.