4 ms·

The recent ChatGPT Images 2.0 has awe-inspiring image editing and composition capabilities. I can totally see what people mean when they say "Photoshop killer".

koonsolo 5mo ago
Same experience on my end. It's crazy what the latest version can pull off, with reference images, text, etc.
- rimliu 5mo ago
  so both of you do not understand what Photoshop is really used for.
  - ACCount37 5mo ago
    Do you? I somehow doubt it. Can you provide an example of something "Photoshop is really used for" that ChatGPT Images 2.0 is completely incapable of doing?
    - arm32 5mo ago
      - Keeping a human face exactly the same while removing imperfections and being able to make small, purely non-destructive edits in series. - Export in very high quality without estimative upscaling.
      - ACCount37 5mo ago
        Are you up to date on frontier capabilities? ChatGPT's image pipeline used to completely suck at sequential edits - i.e. "compounding piss filter" and more. The new Images 2.0 one is a generational leap over that. Have you tried it? Do you still have issues? Not sure what top export resolution is in the system now, but they did bump it up some. What remains is no true native CMYK? But that is a somewhat niche capability. Not sure if you can partially compensate in "pre" by prompting for print-ready images to keep gamuts constrained and get reasonable outcomes from a naive RGB->CMYK conversion.
  - fragmede 5mo ago
    So tell us instead of being vague And condescending?
  - koonsolo 5mo ago
    I used to hire people to do photoshop for me. So yes, I kind of get the idea. I don't hire them anymore. Can you guess why?
vunderba 5mo ago
Yup. You still have to do some manual work if you want to create something like a comic and make sure each panel is internally consistent with the world. Heck, even OpenAI’s own cookbook demonstration of the dog comic was a complete mess. I ended up having to make a lot of manual edits just to fix it. I also think most people fall into the Fiverr camp: they need something that’s “good enough,” and don't have a particularly critical eye for detail - in which case even one-shotting with the latest models, gpt-image-2 or nb2 can get you there.

koonsolo 5mo ago

Same experience on my end. It's crazy what the latest version can pull off, with reference images, text, etc.

rimliu 5mo ago
so both of you do not understand what Photoshop is really used for.
- ACCount37 5mo ago
  Do you? I somehow doubt it. Can you provide an example of something "Photoshop is really used for" that ChatGPT Images 2.0 is completely incapable of doing?
  - arm32 5mo ago
    - Keeping a human face exactly the same while removing imperfections and being able to make small, purely non-destructive edits in series. - Export in very high quality without estimative upscaling.
    - ACCount37 5mo ago
      Are you up to date on frontier capabilities? ChatGPT's image pipeline used to completely suck at sequential edits - i.e. "compounding piss filter" and more. The new Images 2.0 one is a generational leap over that. Have you tried it? Do you still have issues? Not sure what top export resolution is in the system now, but they did bump it up some. What remains is no true native CMYK? But that is a somewhat niche capability. Not sure if you can partially compensate in "pre" by prompting for print-ready images to keep gamuts constrained and get reasonable outcomes from a naive RGB->CMYK conversion.
- fragmede 5mo ago
  So tell us instead of being vague And condescending?
- koonsolo 5mo ago
  I used to hire people to do photoshop for me. So yes, I kind of get the idea. I don't hire them anymore. Can you guess why?

vunderba 5mo ago

Yup. You still have to do some manual work if you want to create something like a comic and make sure each panel is internally consistent with the world. Heck, even OpenAI’s own cookbook demonstration of the dog comic was a complete mess. I ended up having to make a lot of manual edits just to fix it. I also think most people fall into the Fiverr camp: they need something that’s “good enough,” and don't have a particularly critical eye for detail - in which case even one-shotting with the latest models, gpt-image-2 or nb2 can get you there.