Intro
Looks like new v2.5 models from OpenAi with price tag (8-69⚡for 1024px) is an answer to Nano Banana 2 lite (44⚡for 1024px) and Meta Muse (13⚡for 1600px). There is no major image quality improvements. So that is mostly about price.
Price
AR 1:1 is actually more expensive that 3:2 and 2:3
GPT Image 2.5
1024x1024 high - 69 ⚡(7¢)
1536x1024 high - 54 ⚡
1024x1536 high - 54 ⚡
1024px medium - 18 ⚡(2¢)
1536x1024 medium - 14 ⚡
1024x1536 medium- 14 ⚡
1024px low - 8 ⚡(1¢)
1536x1024 low - 7⚡
1024x1536 low - 7⚡
GPT Image 2
1024px high - 287 ⚡(28¢)
1536x1024 high - 222⚡
1024x1536 high - 222⚡
1024px medium - 78 ⚡(8¢)
1536x1024 medium - 65⚡
1024x1536 medium- 65⚡
1024px low - 14 ⚡
1536x1024 low - 14⚡
1024x1536 low - 14⚡
GPT Image 1 and 1.5
1024px high - 375 ⚡(37¢)
1024px medium - 100 ⚡(10¢)
1024px low - 25 ⚡
listed price not includes tips.
Quality
Links to test v3 results
NY 1971 Image
low
medium
high
pareidolia image
low
medium
high
Illustration image
low
medium
high
Solar system on different model versions
compare on solar system (high quality for all version) 3:2 to save some buzz
v2.5 (54 ⚡)
v2.0 (222 ⚡)
v1.5 (375 ⚡)
v.1 (375 ⚡)
prompt is 9D from test v2
Censorship
it is exist. It is not overcensored like som Microsoft or Ideogram models, and also the censorship is quite random.
Mature content triggers
some random definition of mature content by CivitAi. for image like
OpenAi content moderation
OpenAi content filtrated image example
The star wars themed "Ai needs more datacenters" meme (will not post it here, so I got another ToS violation) got filtered on high and medium but generated on low quality settings.
Test v2 - censorship
hmm.. My account is still not muted after couple of "ToS violations" during this tests generation.. so we proceed to test v2 (but test v1 will definitely got me muted, so not today). first of all Now I got more images moderated that it was before
GPT Image 2.0 high
https://wiki.liutyi.info/display/AI/ChatGPT+Images+2.0+high+test+v2
GPT Image 2.5 high
https://wiki.liutyi.info/display/AI/ChatGPT+Images+2.5+high+test+v2
it was 1/40 image filtered before, and now it is 4/40
let's check some image difference
Test v2 - quality
Critic v2.0
Critic v2.5
images a bit different but same level. same order of years issue on the background.
ANNA v2.0
ANNA v2.5
bit more realistic in v2.5, maybe.
Gangster duck v2.0
Gangster duck v2.5
I like both. both are good.
Mirror trap v2.0
Mirror trap v2.5
2.5 a bit better, but still not perfect
Links
see the previous article on v2.0 - chatgpt-images-20-test
see the model page on CivitAi - chatgpt-images
Posts with images from test v3 and v2
(if you need to see the prompts of images used in the article)