Boxes info

[YoungLer]

Hi,
thank for the great work!
May I ask if you have provided the boxes used to calculate Regional PuLID and FLUX-Kontext in Table 2? How can I obtain these boxes?
Additionally, I am wondering how you implemented multi-human generation using FLUX-Kontext, which only supports single-subject generation?
Thanks for your great work again~

[sborse3]

Hi @YoungLer, thanks for your interest in our work!

Regional-PuLID needs non-overlapping boxes, which we curate by dividing the image into N(number of people) separate portions vertically. For 5 people, this looks like:
(0,0),(205,1024)
(205,0),(410,1024)
:
:
(1020,0),(1024,1024)

For Flux-Kontext, results in the MHTB paper are preliminary, but the 'boxes' are for face locations, and the faces are pasted on an input image. We simply make a grid of all the faces, paste it on a single image and input this to Kontext. It doesn't work too well for >3 people, and there is scope for improvement. In related works, Ar2Can(https://arxiv.org/pdf/2511.22690) and Canvas-to-Image(https://arxiv.org/abs/2511.21691) study this kind of method in greater detail.

Let me know if you have further questions!

[YoungLer]

Wow!
THANK for your quick and detailed response. This is a solid work!

I have two other questions:

1. Did you use the "identifier" in multihuman_stetata.json as the regional prompts for Regional PuLID?
2. How to directly utilize LoRA to achieve multi-human generation? Directly using multiple LoRAs will cause blending, which is the problem that Lorashop wants to solve.

THANK for your response~