失敗モード：良い画像を改善に送る

This chapter examines a mistake by the image router. The router decides whether an input image should be enhanced or kept as it is.

Mistaking a good image for one that needs improvement

In the presentation, an already high-quality image of a cheeseburger and fries is used as the example. The image should have been kept as it was. However, the router incorrectly judged its technical quality to be below the bar and sent it down the enhancement path. This is a router false positive: it judged that an image needed enhancement when it did not.

Figure: The cheeseburger-and-fries image is given a routing assessment that its technical quality is below the bar.

The slide places a food image beside the assessment "Technical: Below bar." The other checks pass, while the example is shown as having proceeded to over-processing. The example described in the presentation can be checked at the relevant point in the YouTube video (9:51).

Two problems caused by unnecessary enhancement

First, the system uses compute without an expected quality improvement. Because the image is already good, the quality lift from enhancement is close to zero. Running the editing model still consumes inference compute, money, and sometimes latency. In other words, the system pays for processing that provides no improvement benefit.

Second, the image may become worse. An editing model can add changes that were not in the original image or damage its natural appearance. Unnecessarily processing a high-quality image may therefore do more than fail to improve it: it may reduce the original quality or trustworthiness. Enhancement is not always the safe choice.

The problem happens before editing, not after it

It is important to separate a router error from an editing-model failure. In this example, the first problem is not that the editing model failed to improve the image. The problem is that the system sent an image that did not need improvement to the enhancement path in the first place. Before examining the editing model's ability, we need to evaluate whether the router selects the right images.

Selective enhancement and precision and recall

This failure is connected to the purpose of selective enhancement. The system avoids unnecessary compute and degradation by sending only images that are likely to benefit, rather than enhancing every image automatically. When many good images that should not be enhanced are sent through, the router's precision falls. Precision measures the share of images sent for enhancement that were genuinely worth sending.

By contrast, leaving a poor-quality image unchanged because the system missed it is a recall problem. Recall measures how many of the images that should be enhanced the router successfully catches. If the system focuses only on reducing false positives, it may miss images that really should be enhanced. Router evaluation therefore checks both unnecessary enhancements and misses.

Note: The audio transcript renders the assessment name unclearly, in a form like "technical is low ball." The reliable claim is that the agent assessed the technical quality as low. The presentation does not provide the exact assessment criteria or threshold.

What this example teaches

Sending a good image for enhancement may look like a small classification error. In production, however, it creates both unnecessary compute cost and a risk of image degradation. The system must therefore evaluate not only image-editing quality but also the router's decision about whether enhancement should happen, as a separate target.