In this chapter, we check things that cannot be judged from the image alone together with its description or metadata. Even if an image looks natural, that does not always mean it is safe to publish.
An image lets us check colors, shapes, and arrangement. However, pixel inspection alone cannot always verify a textual claim such as “there are eight wontons.” Even if the dish looks natural, its visible item count may not be possible to confirm accurately.
That is why we use multimodal evaluation. It examines the image together with a description or metadata and checks whether the two kinds of information agree. Instead of evaluating only the image, it asks, “Does this image match this description?”
The slide shows an example in which the image and a description saying “8Pcs (8 pieces)” are checked together. The count in the description is part of the evaluation, not just the visual appearance. However, the slide itself does not explicitly show that the evaluation was uncertain or that the case was rejected in production.
According to the speakers, looking at the image and output does not establish that eight wontons are really present. The system therefore treats it as a case where it lacks confidence and rejects it in production instead of accepting it. In other words, it does not add nonexistent wontons through editing just to match the description.
This is a “do not publish on failure” decision that protects safe routing and faithfulness. It is important to retain an “unsure” state rather than using only “pass” and “fail.” If an uncertain case is marked as a pass, incorrect content or an image that does not match its description may be published.
The important distinction is that insufficient evidence is different from a proven content mismatch. For example, if it is clearly visible that an image contains only six items, we can identify a mismatch with a description saying eight. In this chapter’s example, however, we cannot confirm that eight are present, so we reject it without confidence. This does not assert that a visual failure has been proven.
Note: The speakers do not explain which metadata field supplied the claim of eight items or how confidence was calculated. The reliable point is the operating policy: combine the image and text, and stop publication in production when the evidence is insufficient.
In short, a multimodal system must do more than create a natural-looking image. It must check the relationship among the image, its description, and its metadata, and it must not force editing or publication when the judgment is uncertain. This conservative decision supports faithfulness and production safety.