It's truly fascinating to see how photography, CGI, and Photoshop are being "replaced" in these videos. I understand the appeal of this technology, and Nano Banana is undoubtedly a step forward, but we're talking about 1024x1024 images and 80% reproduced likeness here. While that might suffice for YouTube videos or profile pictures, in the real world of high-end production, we work with images up to 400MP daily. I am very adamant taking resolution as a single quality factor because usually I get ESRGAN upscaled images as a counterpoint, look the resolution is here, what people are missing is the fidelity of say a phase one or Hasselblad.
You can achieve these higher resolutions, fidelity and close to 100% reproduction of objects with considerable effort in generative AI right now, it's even a very enjoyable experience to do so, but not with a closed model like Nano Banana. When Google or other major image model providers give us proper fine tuning and custom inference, particularly for applications like upscaling, then we can have a serious conversation.
At present, while I understand the fascination, everyone should take a step back and consider what we're actually trying to replace. We are stuck at these low inference resolutions since SDXL. For genuine high-end commercial applications, we're not there yet, and we won't be without another significant leap in GPU technology.
fallinditch · 2h ago
One of the things I enjoy most about photography is using old film cameras. They are just tools but to me they are suffused with mystery and magic, and some are simply beautiful [1]. It's like they have a power that is missing from our convenient digital world, and in the act of being used they shape and mediate our creative choices.
400M pix shot of a genetic banana replaced by a 400M pix synthetic shot of a reputed banana, multiplied by Youtube channel mansplaining the finer points of reputed banana synthesis.
The only thing missing is any interest whatsoever in actual bananas!
Take career notes kids, as the AI bandwagon is determined to exhaust the entire world and stuff its remains along with cases of Dom Pérignon down into bomb shelter made from an abandoned nuclear missile silo for the benefit of a VC and his "Tech Support" bros by reputing everything the same way it's reputed the banana.
You can achieve these higher resolutions, fidelity and close to 100% reproduction of objects with considerable effort in generative AI right now, it's even a very enjoyable experience to do so, but not with a closed model like Nano Banana. When Google or other major image model providers give us proper fine tuning and custom inference, particularly for applications like upscaling, then we can have a serious conversation.
At present, while I understand the fascination, everyone should take a step back and consider what we're actually trying to replace. We are stuck at these low inference resolutions since SDXL. For genuine high-end commercial applications, we're not there yet, and we won't be without another significant leap in GPU technology.
[1] https://www.mes-appareils-photos.fr/Zeiss-Ikon-Nettar-517-16...
400M pix shot of a genetic banana replaced by a 400M pix synthetic shot of a reputed banana, multiplied by Youtube channel mansplaining the finer points of reputed banana synthesis.
The only thing missing is any interest whatsoever in actual bananas!
Take career notes kids, as the AI bandwagon is determined to exhaust the entire world and stuff its remains along with cases of Dom Pérignon down into bomb shelter made from an abandoned nuclear missile silo for the benefit of a VC and his "Tech Support" bros by reputing everything the same way it's reputed the banana.