GPT-Image 2.5 just solved character consistency. Here's the back-to-back comparison along older models:

Lowkey gpt image 1 kinda cooked

cool, but I think we can consider gpt image 1 as true ai art

2 years ago. You guys are such stupid assholes.

GPTIMG1

what if they just added some logic that says if "don't change the image" is the prompt, skip the model and just return the image.

look like jpegxl, lossy file format with degradation in recompression !

Oh hell yeah they solved the texture problem with 2 didn’t realize that

Didn't NB PRO solve it ?

Same prompt, same character, way less drift means massive creative control.

character consistency finally solved??

Lol why was replica hard for it

back-to-back samples miss the hard case: identity surviving heavy occlusion

角色一致性这关要是稳了，做连续画面舒服多了

Give me a higgsfield subscription and ill break it open

Gosh it looks like the mandela catalogue

It looks like tom lee from lookism (GPT 1)

Can we stop calling one clean comparison “solved”? It looks better. That’s enough of a claim.

Wow

not best of course, but much better.

Why it can't create good Manga or comics it is really bad at that??? I wish you guys will fix it

gpt 1 straight made you into Dracula

1.0 was dark side

This is changing how people make creatives

LMAO the noise issue with images 2

Very cool

Promising result. The harder test is keeping the same person stable across many scenes, edits, clothes, injuries, and lighting changes as the visual history grows.

👀

W

character consistency is key for storytelling, especially in games and films

Does the identity anchor hold up when switching to extreme camera angles or dynamic lighting? Usually multi-turn generation starts drifting once pose variance gets too aggressive.