Component-wise ablation of the S-GDR conditioning branches

Establish the individual contributions of the Qwen2-VL semantic prompt, ControlNet, and IP-Adapter conditioning branches to the performance of the Semantically-Guided Domain Randomization pipeline by conducting a leave-one-component-out ablation.

Background

The comparison between guided and random prompts isolates semantic grounding only while ControlNet, IP-Adapter, SDXL, and mask-based composition remain active in both configurations. Thus, the current experiment cannot determine how much performance is attributable to each conditioning component individually. The paper explicitly identifies a leave-one-component-out experiment as unresolved.

References

Similarly, the effect of alternative captioners (e.g., LLaVA) and of prompt-engineering choices is not resolved by the current data.

— Semantically-Guided Domain Randomization for Industrial Object Detection in Low-Image-Budget Regimes  (2609.26505 - Araya-Martinez et al., 22 Sep 2026) in Section 5.1, What the Prompt Comparison Can and Cannot Attribute; Section 6, Limitations and Potentials, item 4