This chapter explains what a router that chooses the next processing path receives and how it makes its decision. In the presentation, the system combines an image, a text description, and metadata, structures that information, and then decides whether to enhance or skip the image.
ルーターが受け取る三つの情報
The three kinds of information the router receives
The information here is not only an image. Using several kinds of information together, such as images and text, is called multimodal. This router looks at the following three inputs together.
The router looks at the three inputs and first asks for a structured description. The idea is to turn the inputs into an intermediate representation that is easier to use in later evaluation, rather than making a decision from free-form text.
Supplement: When the inputs are combined into one structured representation, the router can check what the input contains and then evaluate it against criteria. This separates the stage that describes the input from the stage that selects a processing path. However, the presentation does not disclose the schema or specific fields of the structured data.
流れを整理すると、次のようになります。
To summarize the sequence:
画像、テキスト、メタデータを参照します。
それらをもとに、構造化された画像の説明を作ります。
その表現をルーブリック（rubric）で評価します。
評価結果から、強化（enhance）するか、スキップ（skip）するかを決めます。
Refer to the image, text, and metadata.
Use them to create a structured description of the image.
Evaluate that representation with a rubric.
Use the evaluation result to decide whether to enhance or skip.
Figure: A three-stage routing flow that refers to image and text inputs, evaluates them with a rubric, and decides whether to enhance or skip.
このスライドは、発表で説明された順序を直接示しています。左側で画像とテキストを参照し、中央で結果をルーブリックに照らして「基準未達（Below bar）」または「合格（Pass）」として評価し、右側で強化またはスキップを決めます。画像の例では、Technical と Dish match が基準未達で、Composition、Content、Policy は合格になっています。これは、見た目だけでなく、料理の説明との一致やポリシーも判断材料になることを示します。
This slide directly shows the order described in the presentation. On the left, the system refers to the image and text. In the center, it evaluates the result against a rubric as “Below bar” or “Pass.” On the right, it decides whether to enhance or skip. In the image example, Technical and Dish match are below bar, while Composition, Content, and Policy pass. This shows that the decision uses not only appearance but also agreement with the food description and policy.
A rubric is a set of explicit criteria for making a decision. The router evaluates the structured description against these criteria and treats each item as pass or below bar. This converts the model's free-form description into an operational decision instead of connecting it directly to a processing path.
Example: Keep the description and evaluation separate. First, structure what appears in the image. Next, evaluate whether that content meets criteria such as technical quality, composition, dish match, and policy. Finally, use the result to choose enhance or skip. These criteria are an explanatory example; the presentation does not claim to show every field in the actual rubric.
「評価」と「処理先の選択」は別です
“Evaluation” and “choosing a processing path” are different
This flow requires us to separate two similar problems. One is whether the input can be described and evaluated correctly. The other is whether the correct processing path can be selected. The router's job is the latter. Even if the router chooses enhancement, the later edit is not necessarily good. Conversely, choosing skip is not an editing failure; it is a decision not to edit.
段階
主な役割
出力
入力の説明
画像、テキスト、メタデータをまとめて内容を表す
構造化された説明
ルーブリック評価
明示的な基準に照らして確認する
合格／基準未達
ルーティング
評価結果に応じて処理先を選ぶ
強化／スキップ
Stage
Main role
Output
Input description
Combine the image, text, and metadata to represent the content
Structured description
Rubric evaluation
Check the representation against explicit criteria
This binary choice supports selective quality improvement. Instead of sending every image through an expensive editing process, the system sends only images worth enhancing. Unnecessary processing increases compute cost. Also, editing an image that is already good may produce no quality gain and may degrade the original image.
ルーティングの正しさ：その画像を強化側またはスキップ側へ送る判断が適切か。
編集の品質：強化側へ送った後の出力が、元画像より良く、内容にも忠実か。
Routing correctness: Whether it is appropriate to send the image to the enhance or skip path.
Editing quality: Whether the output after enhancement is better than the original and remains faithful to its content.
These are separate evaluations. Even when the router makes the right decision, the later edit or QA (quality assurance) can fail. Conversely, even if the editing model performs well, sending an image that does not need enhancement can create unnecessary cost or degradation risk.
The presentation shows a design that uses image, text, and metadata, evaluates a structured description with a rubric, and then chooses a route. However, it does not show the schema of the structured representation, all rubric fields, or the thresholds separating pass from below bar. Therefore, we should not infer a specific JSON format or numeric thresholds.
まとめ
Summary
ルーターは画像だけでなく、テキストの説明とメタデータも使います。
まず入力を構造化された中間表現にし、次にルーブリックで評価します。
合格／基準未達という評価を、強化／スキップという運用上の判断につなげます。
ルーティングの正しさと、後の編集結果の品質は、別々に評価する必要があります。
The router uses an image, a text description, and metadata.
It first creates a structured intermediate representation and then evaluates it with a rubric.
It turns pass / below-bar results into the operational choice of enhance / skip.
Routing correctness and the quality of the later edit must be evaluated separately.