Disentangled Object Sampling for multi-object generation

[cyf23]

Hi authors,

Thanks for sharing this amazing work and the codebase!

I am currently trying to reproduce the multi-object generation results. In the paper, Algorithm 1 (Disentangled Object Sampling) is proposed to solve attribute leakage by using masked latent blending ($x_{t+1} = (1 - M_i)x_{t+1} + M_i x^i_{t+1}$) and localized text conditioning.

However, when reviewing app.py and pipelines/pipeline_scenedesigner.py, it appears that the current pipeline acts as a standard global SD3 ControlNet. The input seems to be a single global prompt and a global full_scene.exr condition, without the inner loop for local object generation and latent mask blending mentioned in Algorithm 1.

Could you please clarify if:

1. I missed a specific branch or script (e.g., a multi-object inference script) that contains the Algorithm 1 logic?
2. Or is the Disentangled Object Sampling code planned for a future update?

Any guidance would be greatly appreciated. Thanks again for your great contribution!