I'd be interested to know the parameters used, especially prompt_strength
The correspondence to the original image is not especially high: the Leisure Suit Larry image, for example, enhances the original colours of the sea in a nicely realistic way, but all the foreground detail is essentially reinvented from scratch, including some very obvious omissions. In some of them the changes to perspective and more lifelike skull/canyons etc might improve on the original image, but it also flips even pretty basic stuff like which shoulder the woman's hand is placed on (and yes, once you look at that hand closely, the fingers SD has had to add in are all wrong...)
Ideally for this sort of use case you'd want high fidelity to the geometry of the original image but less fidelity to the palette (use more than 256 colours and naturalistic or artistic textures rather than lines and pixel dithering), but I'm not sure SD can manage that at the moment
The correspondence to the original image is not especially high: the Leisure Suit Larry image, for example, enhances the original colours of the sea in a nicely realistic way, but all the foreground detail is essentially reinvented from scratch, including some very obvious omissions. In some of them the changes to perspective and more lifelike skull/canyons etc might improve on the original image, but it also flips even pretty basic stuff like which shoulder the woman's hand is placed on (and yes, once you look at that hand closely, the fingers SD has had to add in are all wrong...)
Ideally for this sort of use case you'd want high fidelity to the geometry of the original image but less fidelity to the palette (use more than 256 colours and naturalistic or artistic textures rather than lines and pixel dithering), but I'm not sure SD can manage that at the moment