Thank you for open-sourcing UniWorld — it's an impressive and elegant design!

I have a question regarding your two-stage training procedure. In Section 3.2 of the paper, you mention that Stage 1 is used to align the VLM features with the FLUX text branch via a frozen setup, and that skipping proper alignment (e.g., introducing T5 features too early) tends to cause the model to converge to poor local minima.

I’m curious:
Have you tried skipping Stage 1 and training the entire architecture directly from Stage 2 (i.e., unfreezing all parameters from the start)? Would such a strategy lead to divergence or collapse (e.g., NaNs, trivial solutions, or failure to use SigLIP features effectively)? Or would the model still converge, albeit to a suboptimal solution?

Understanding this would be very helpful for ablation or simplification purposes.

Thanks again for your great work!