Hierarchical approach
#3
by stablellama - opened
Have you considered an hierarchical approach?
I.e. for image generation denoinse a much lower res image up to a middle step count (e.g. 7 when the target is 10), double the height and width (perhaps with adding some noise at the same time) and then start the denoising again (perhaps at step 3 or so) and do that a few times till the target dimension is reached?
I guess that a pixel space model, especially one that can also upscale, could do that very well and this would cut down generation time while also allowing huge output dimensions.
(The numbers I have sated should give the idea, the best numbers must be worked out)