Question about control branch

[heishuini]

I noticed that there are two noisy_lq_latents in the code, which are ultimately fed into the transformer by concatenating them in sequence. This approach results in higher computational costs. May I ask:

1. Compared to this, would channel-wise concatenation be better?
2. If only one noisy_lq_latent is introduced, restoration can still be achieved, but the performance seems worse than your sequence concatenation method. Is my understanding correct?
   Thank you for your reply!

[double8fun]

1. Channel-wise concatenation should work, but it may require model structure modification, and in-context manner is proved to be effective in previous work (e.g. Flux-Kontext).
2. Yes, as shown in the ablation study.