[bug]: ZImage-Turbo regional guidance not working as expected

[lstein]

Is there an existing issue for this problem?

• I have searched the existing issues

Install method

Invoke's Launcher

Operating system

Linux

GPU vendor

Nvidia (CUDA)

GPU model

RTX 5060 Ti

GPU VRAM

16GB

Version number

Main branch commit 63ff383

Browser

Chrome

System Information

ZImage-turbo regional guidance is not working in the way I expect. With other generation models, when I define two or more regional guidance layers, they end up being composited together to form a single unified image. With ZImage-turbo, I end up with two independent images.

What happened

Here is an example. I have defined two regional prompt guidance areas and an overall prompt of "two women." The first regional guidance area specifies a woman in a business suit. The second specifies a woman in a peasant dress.

Screenshot of the canvas:

With flux, I get a single unified image of two women positioned in the expected way:

With ZImage-turbo, it looks like this instead:

Is this the expected behavior?

What you expected to happen

I expect a unified image.

How to reproduce the problem

No response

Additional context

No response

Discord username

No response

[Pfannkuchensack]

Overly strict isolation: The regions are so tightly isolated that the model treats them as completely separate generation tasks.

Missing overlap zone: Unlike Flux, there is no explicit handling of overlap zones between regions.

Model architecture difference: ZImage-Turbo uses a different architecture than Flux (unified sequence instead of double/single stream), which may require different attention strategies.

[lstein]

I've just given the original implementation (before #8718 ) a try with the same prompts but overlapping regions, and it already addresses the issue pretty well. Seems like I haven't been using regional prompting correctly all this time!

I'll try now with the PR.