Write a PREreview

Pixel-level annotation remains a major bottleneck for semantic segmentation, motivating methods that synthesize image-label pairs directly from generative models. Prior synthetic dataset generators typically obtain pseudo-labels from cross-attention maps or learned decoders over generative features; however, recent text-to-image (T2I) models increasingly use multimodal diffusion transformers (MM-DiTs), where concept localization is no longer exposed through a single cross-attention pathway but instead distributed across many layers and attention heads. Existing MM-DiT localization methods address this by aggregating saliency across heads, but we observe that this averaging can dilute clean target localizers due to attention head heterogeneity. We introduce HEADHUNTER, a training-free, open-vocabulary segmentation framework that uses aggregate concept saliency as a self-guided proxy to select the single attention head that best localizes a queried textual concept, yielding cleaner segmentation masks. We then use HEADHUNTER to turn target classes into training data automatically: a large language model (LLM) diversifies prompts, an MM-DiT generates images, HEADHUNTER produces pseudo-labels, and a vision language model (VLM) verifies each image-mask pair before acceptance. On PASCAL VOC2012, HEADHUNTER achieves 79.2 mIoU in zero-shot single-class segmentation, outperforming MM-DiT head aggregation. Using only our generated image-label pairs, DeepLabV3 with a ResNet-101 backbone reaches 64.9 mIoU on VOC2012 validation, matching or outperforming comparable synthetic dataset generators and showing that the proposed pipeline produces effective dense labels without human intervention. Code: https://github.com/Rohanl2298/HEADHUNTER.