Users report GPT-image-2 outputs remain hard to control for art generation because of limitations in the underlying GPT model, unlike Midjourney which accepts detailed prompts verbatim
Midjourney is the sole generator accepting detailed prompts verbatim.
What I mean concretely is that with the "same" input and massive guidance GPT-Image-2 (extended thinking) produces something like left image and Z-Image something like the right. I know what I'd use for a cover of an "AAA" game for a Global South audience. But… it's not art.
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexGpt-image-2 is surprisingly bad for art, or rather, hard to make usable for it. and the problem is almost certainly GPT
if you use the API, this comparison is also possible "Intention 9: Consuming the God"
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexWhat I mean concretely is that with the "same" input and massive guidance GPT-Image-2 (extended thinking) produces something like left image and Z-Image something like the right. I know what I'd use for a cover of an "AAA" game for a Global South audience. But… it's not art.
my honest reaction to this information
Andrew Curran@AndrewCurran_@teortaxesTex You get different results with Pro, which supports this.
@teortaxesTex I don't think anyone except midjourney lets them go in verbatim anymore.
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexmy honest reaction to this information