Transformer model for DLSS is honestly overrated atm, most of the notable differences so far is because it looks sharper over the CNN model. But CNN by default is just generally blurry due to a lack of sharpening and using a sharpening filter with it automatically brings out more detail to the point where it is on par with the transformer model. I've been testing it for over 2 weeks now and for a lot of games, not only does the CNN + sharpening give better performance, but it also provides a slightly more stable image a lot of the time.
The transformer model tends to produce a lot of localised shimmering on objects and textures that isn't present with the CNN model. Even with ghosting reduction which is the main touted benefit of the transformer model, while there is a general reduction, it's still present with some fast-moving objects and again there are scenarios where things that shouldn't be ghosting are displaying noticeable ghosting. E.g. moving the camera can sometimes cause static objects/textures to ghost where they wouldn't have with the CNN model.
Most of the praise I have seen for it comes from people who complain that DLSS3 was too blurry while seemingly not understanding that they should have been using a sharpening filter, especially after Nvidia disabled inbuilt sharpening with ~DLSS 2.5. So really I think if FSR4 can match DLSS3 image quality and adds a bit of sharpening by default, the differences to DLSS4 will not be considerable at all.
2 gens in a row, not a great look.
