Do you actually use 4 dataset in MMAD in your training set?

[tzjtatata]

I am curious about the training dataset.
Because there is no way to see the product before under the zero-shot IAD settings. But I see the 4 datasets used in MMAD in Fig. 7.
Do EMIT actually see the images during either SFT or GRPO?

[guanwei49]

Thank you for your question.

Under the zero-shot IAD setting, our method does not see these images during either SFT or GRPO training. As described in Section Object Descriptive Text as Query Image, we replace missing defective images with object descriptive texts. Specifically, the training data consist of (1) descriptive texts of normal images, and (2) simulated anomaly descriptions generated based on expert domain knowledge (defect type, location, and characteristics), rather than real images.

To further verify this design choice, we conducted an ablation study by removing all descriptive text samples from the training data. While these text samples help the model learn more effectively, the results show that our method still outperforms existing approaches even without them, demonstrating that EMIT does not rely on seeing product images beforehand.

Therefore, although Fig. 7 presents four datasets used in MMAD, EMIT does not access these images during training; it instead leverages text-image semantic alignment to enable training without visual samples.

[tzjtatata]

Thank you for reply. I have noticed the ablation. But without text examples, EMIT achieves only 71.07% which does not surpass InternVL3-8B with pure GRPO (71.44%). Is the result something wrong?

[guanwei49]

Thank you for pointing this out. The result is expected and there is no issue with the experiment.

InternVL3-8B with pure GRPO (71.44%) is trained on the full training dataset, which includes both image-based samples and object descriptive text samples. In contrast, the “w/o text examples” setting uses a strictly smaller training set, as all text-only samples are removed. The performance drop (71.07%) therefore reflects the reduced data scale rather than a flaw in the method.
Even without text examples, EMIT remains competitive with GRPO-based baselines, and when text examples are included, EMIT shows a clear and consistent improvement.

This confirms that the descriptive text is beneficial for learning, while EMIT’s overall gains stem from the combined effect of the training strategy and dataset design.

[tzjtatata]

But do the text examples also influence the GRPO?

[guanwei49]

Exactly. The text examples are available to any method and do not provide EMIT with any exclusive advantage. They are part of the training data construction and can be equally used by other GRPO- or SFT-based approaches.