Mitigating performance degradation through additional data and model capacity

Determine whether increasing the available data and model capacity can mitigate the performance degradation observed as the categorical task family becomes broader and more complex in the Food-101 and QuickDraw in-context diffusion experiments.

Background

The experiments on QuickDraw and Food-101 show that performance generally degrades as the number of categorical task distributions increases under fixed per-task data, prompt size, and shared model capacity. The paper attributes this trend to the increasing complexity and heterogeneity of the task family rather than to the theoretical task-sampling rate under a fixed meta-distribution.

The authors qualify this degradation as a resource-limited phenomenon rather than an inherent limitation of the proposed in-context diffusion method. They explicitly leave unresolved whether supplying more data and increasing model capacity would alleviate the observed deterioration.

References

Whether additional data and model capacity can mitigate this degradation is left for future work.

— Transformer-based In-Context Score Operator Learning for Diffusion Models  (2610.13690 - Xu et al., 9 Oct 2026) in Section 4, subsection “Food-101 Dataset”