MiniMax H3 single frame edit excels at character consistency and identity preservation. Do we really need character LoRAs anymore?

Basically MiniMax H3:

• Has great prompt following
• Respects character identity (art style and design) from the input images
• Can generate single frame images

So what’s the point of training a character model if you can just give MiniMax H3 character design reference and generate any scene? The only two reasons I see is
1) you want to use style known by your target base model to generate image of the learned character.
2) you want to apply some concepts known by the target base model only.

What do you think?

Here’s the workflow I used for the images.