The editing model has a very ordinary effect
The T2I model performs great, but editing the model falls far short, especially in terms of the consistency of the edited characters, which has almost no relation to the reference image. Is this a problem with the model
"The official limitation section is unusually useful. They say Boogu still has gaps versus strong closed systems in world knowledge, real brands/people/landmarks, and complex contextual understanding. They also say image-to-image consistency is not stable enough for strict preservation, that it trails Seedream 5.0 and Nano Banana Pro in some in-context editing scenarios, that long/dense text can still typo or drift, and that complex poses/multi-person interactions can still break hands/limbs/body structure." (GPT summarizing from the -HF page. Also that it is worked on (0.1)
Although "it has "some gaps" compared with nanabanana2 borders on boasting.
so yes. it's the model's own weakness. This model came out of the blue.... looking foward to it, seems to get ok reactions apart from the editing . Lots of small-ish models recently. Ernie was just crap, ideogram is nice but complicated and hidreamo1 blurry. Let's see where this apparently completely new one goes.