制限付きの改善・QAループ

This chapter looks at the enhancement and quality-assurance (QA) flow after the router sends an image to the enhancement path. The point is not to edit once and finish. The system uses the QA result to correct the edit. However, it does not repeat forever until it succeeds.

1. Enhancement proceeds in three stages

The presentation divides the enhancement process into the following three stages. The initial inputs support the decisions in the later stages.

1. Create an image-specific editing prompt from the image description and routing directives.
2. Edit the image using that prompt.
3. Check the edited result from multiple perspectives with QA.

In the first stage, the system uses the image description and the directives produced by the router. It does not create a general prompt that performs the same action on every image. It creates an image-specific prompt suited to what this image needs to improve. For example, the editing instruction differs for an image that needs to be brightened and one that needs its plating corrected. The presentation does not show the exact prompt format.

Next, the editing agent enhances the image. This enhancement is not the unconstrained creation of a new image unrelated to the original. It uses the description and routing directives as its basis and performs the edits needed for the target image. Therefore, editing quality is not determined only by how attractive the result looks. It is also important whether the original dish and information remain intact and whether the improvement is natural.

The QA gate in the third stage does not look at only one aesthetic score. The perspectives named in the presentation include plating, faithfulness to the original image, and colors. It checks separately whether the plating is appropriate, whether content absent from the original was added, and whether the colors changed unnaturally. The enhancement result moves forward when it satisfies these multiple conditions.

2. Send QA failures back to editing

When QA finds a problem, the result is not merely recorded. QA sends feedback about the problem back to the editing side. The editing side combines the initial inputs with the new QA feedback and enhances the image again. In other words, this is not a one-way process of “edit and check”; it is a closed loop in which the check result is used for the next edit.

The order of this loop is as follows.

1. Create an image-specific prompt from the image description and routing directives.
2. The editing agent generates and edits the image.
3. QA checks plating, faithfulness, colors, and other dimensions.
4. If it passes, move the result forward. If it fails, send the QA feedback back to enhancement together with the original inputs.

This feedback is not meant to discard the original prompt and start a different objective. It is used to correct the discovered problem while preserving the description and instructions for the original image. This boundary helps distinguish the QA correction loop from unconstrained generation that remakes the image into different content.

Figure 1: The “Generation” diagram shows three stages: create an image-specific prompt from the description and routing directives, enhance the image, and check it from multiple perspectives with QA. A result that fails QA returns to retry with QA feedback. See the video around 13:37 for details.

3. Set a limit of K iterations

This QA loop has a maximum of K iterations. K is a symbol for the number of times enhancement may be tried. The presentation does not give the specific value of K or the thresholds for each QA condition. Therefore, no particular count or pass rate can be inferred from it.

If QA passes before the limit is reached, the passing enhancement result proceeds to publication. If it still fails after K tries, the system does not force out a bad result. It treats the image as not enhanced and accepts a reduction in coverage, meaning a narrower range of images can be improved. This choice prioritizes not publishing a problematic image over increasing the number of images that receive enhancement.

The presentation does not quantify this “coverage impact.” It can be understood as a lower share of images that can be safely passed, but its exact definition and value are unknown. The important point is to provide a bounded failure path and not treat a result whose quality cannot be confirmed as a success.

Figure 2: This diagram connects prompt generation, image enhancement, and a QA gate, then sends a failed check back into the loop together with QA feedback. It does not explicitly show the K-iteration limit, publication branch, or coverage reduction. Those are operational rules explained in the presentation. See the video around 14:06 for details.

4. See the decision through a concrete example

Example (a hypothetical example for explanation): A food photo may become easier to see if its colors are made slightly brighter. However, if the edit disrupts the plating or adds an ingredient that was not in the original image, it cannot pass QA even if it looks brighter. QA checks not only whether the image became more attractive, but also whether it remains faithful to the original content and whether the plating and colors are appropriate.

If the first edit fails QA in this hypothetical example, the feedback becomes material for the next edit. The editing side keeps the original description and instructions while correcting the plating or added-content problem. If the conditions are still not met within K iterations, the system does not force the result into publication. This example does not guarantee that more iterations will always succeed. More iterations can also increase compute cost, latency, and the risk of over-editing.

5. Keep this QA separate from final publication QA

The QA in this chapter is an iterative gate for correcting the edited result. At a later stage, there is publish-ready QA that checks whether the post-processed result may be published. Therefore, passing this chapter’s QA is not the same as being ready for production publication. The first helps correct the edit; the second acts as a separate defense immediately before publication.

What can be said with confidence from the presentation

The presentation clearly shows that the system creates a dedicated prompt from the image description and routing directives, edits the image, and performs multidimensional QA including plating, faithfulness, and colors. If it fails, the system sends feedback back with the inputs and tries up to K times. If it passes, the result proceeds to publication; if it does not, the system loses coverage by not enhancing it. The value of K, the detailed QA implementation, thresholds, and prompt format are not disclosed.

Summary

1. Enhancement has three stages: create an image-specific prompt, edit the image, and perform multidimensional QA.
2. A QA failure returns to editing as feedback together with the original inputs and creates the next attempt.
3. The loop is limited to K iterations; a passing result moves forward, but a failed result is not forced into publication.
4. To protect faithfulness and safety, the system keeps a non-publication path when a failure occurs instead of pursuing only pass rate or coverage.