This chapter organizes the goals that a multimodal agent for improving food images should pursue at the same time. The presentation is not only about making images look better. It treats authenticity, user trust, marketplace health, safe publishing, continuous learning, and the cost of large-scale operation as one design.
発表では、次の六つの目標が示されます。
The presentation gives the following six goals.
本物らしさと信頼を守る
必要な画像だけを選んで品質を上げる
特定の加盟店を犠牲にせず、マーケット全体を最適化する
安全に公開する
継続的に学習する
大規模に、コスト効率よく動かす
Preserve authenticity and trust
Improve only the images that need improvement
Optimize the marketplace as a whole without sacrificing particular merchants
An enhancement must not move away from the original dish or the merchant's characteristics. It is important that users can feel that the image represents what they will actually receive. If an image looks AI-generated, it may lose trust even if it looks polished. The goal is therefore not to maximize an artificial appearance. It is to make a useful improvement while preserving the connection to the original image.
The system should not edit every image automatically. It should select images for which improvement is worthwhile. Processing an image that is already good can use compute without increasing quality. It can also risk making a good original worse through editing. The agent therefore first examines the image and decides whether to enhance it or keep the original. This design asks "Should we edit?" before it asks "Can we edit?"
This is a teacher-created example. If a photo has no major problem with brightness or composition, the system keeps the original. It sends only photos with problems that appear improvable to editing. Separating these cases reduces unnecessary processing and unnecessary quality loss. The presentation does not specify the criteria or exact thresholds for making this distinction.
3. 特定の加盟店を犠牲にせず、全体を最適化する
3. Optimize the whole marketplace without sacrificing particular merchants
The target of optimization is not one image or one merchant. The system should consider the health of the marketplace as a whole. It should avoid a result that makes one merchant's numbers look better by taking opportunities away from another merchant. The presentation warns against sacrificing particular merchants in this way, using the term cannibalization. In other words, we must not look only at an average quality increase and overlook harm to individual merchants.
例：全体の改善と一部の悪化を分けて見る
Example: distinguish overall improvement from local deterioration
As an additional explanation, even if an overall metric improves, the design goal has not been fully met if results become worse for a particular region or group of merchants. We must also check which merchants or segments changed, not only the overall result. The later part of the presentation connects this idea to segment-specific tuning.
The system should not show an agent's enhancement to users without further checks. It places safety boundaries, or guardrails, that stop problematic results before publication. In the later workflow, stage-level quality checks, logs, human-provided labels, publish-time checks, and rollback to an earlier version when a problem occurs support this goal. This chapter does not specify the details. It establishes that safe publishing is a design requirement from the start.
Production images and user responses change over time. An agent tuned once will not necessarily keep working well forever. The system therefore needs a cycle that collects new examples from real data, compares them with human judgments and operational feedback, and adjusts the necessary parts. Continuous learning here does not necessarily mean retraining model weights. Later, the presentation explains a process that updates settings or prompts, checks them, and then uses the new version.
When the system handles diverse images from around the world, producing one good example is not enough. It must process many inputs reliably while managing compute and cost. This is why selective enhancement is also connected to the cost goal. Sending only necessary images through expensive processing, while producing safe and useful results overall, is a requirement of large-scale operation.
The six goals are not separate checklists. They are multiple objectives that affect one another. If the system pushes visual improvement too strongly, it may lose authenticity or faithfulness to the original. If it stops everything for safety, the scope of quality improvement becomes too narrow. Even if learning is automated, the system is not safe without guardrails and publish-time checks. It must also satisfy the requirements of cost and scale, not only quality.
Therefore, later evals do not measure only one score for "looking good." They check which images should be enhanced, whether the original content is preserved, whether the result can be published safely, whether the marketplace as a whole is harmed, and whether the system operates at an acceptable cost. The design goal is not to remove the agent's flexible judgment. It is to allow useful improvement while keeping boundaries for trust, safety, diversity, and cost.
図：「Our Goals」と題したスライドに、画像エージェントの六つの目標が同時に示されています。
Figure: A slide titled "Our Goals" shows the six goals for the image agent at the same time.
This slide directly shows that there is not just one goal. Authenticity, selective quality improvement, marketplace-wide optimization that preserves merchant diversity, safety, continuous learning, and scalable cost-efficient operation appear together in the same design. The relevant part of the presentation starts at 4:08.