Question on text representation bottleneck vs. method bottleneck

[cswry]

Hi, thanks for releasing this very interesting work.

I really liked the paper’s perspective that few-step T2I generation may be especially sensitive to text representation quality, and that discriminability / disentanglement could be an important bottleneck for MeanFlow in text-to-image settings.

I have one question about how to interpret this claim.

I recently read Few-Step Distillation for Text-to-Image Generation: A Practical Guide, which compares several few-step distillation methods on the same T2I teacher (FLUX.1-lite). Their results seem to show that MeanFlow underperforms sCM quite significantly in the ultra-few-step regime: sCM stays strong at 1/2 steps, while MeanFlow collapses at 1 step and is still much weaker at 2 steps, only becoming competitive around 4 steps.

At the same time, your Figure 4 seems to show that in your final T2I MeanFlow setup, 1-step performance keeps improving with longer training rather than failing completely. My current reading is that Figure 4 mainly shows a convergence issue within your final setup (with the representation/framework fixed), rather than directly isolating whether text representation is the main causal bottleneck.

So I wanted to ask:

1. Is this reading of Figure 4 correct?
2. Do you see text representation as the main bottleneck, or as one important bottleneck among several?
3. Have you done any controlled comparison that keeps the same T2I teacher/backbone and text representation fixed, while comparing different few-step distillation methods, to separate representation bottleneck from method bottleneck?

I’m asking sincerely because I think the paper raises an important perspective, and I’d really like to better understand the scope of the claim.

Thanks again for the great work.