Hey, I find pixelspace models fascinating, and I wanted to know if you have considered using a colorspace that behaves more liberally perceptually, like Oklab.
From personal experience with rdo (rate-distortion optimization) for images, measuring the distortion in oklab makes a noticeable difference. So my thought here was that there might be some free performance to gain, buy diffusing in oklab/using loss in oklab.
Hey, I find pixelspace models fascinating, and I wanted to know if you have considered using a colorspace that behaves more liberally perceptually, like Oklab.
From personal experience with rdo (rate-distortion optimization) for images, measuring the distortion in oklab makes a noticeable difference. So my thought here was that there might be some free performance to gain, buy diffusing in oklab/using loss in oklab.