Failure mode: a recall miss and cross-modal hallucination risk

An image can look usable and still be wrong for the item it represents. This chapter shows a routing failure in which the system accepts an image that should have been caught. The failure is especially dangerous because the text description can then push the image editor to invent content that is not present in the source.

Source focus (10:14–10:45): The presenters show an image with six chicken wings paired with a description of eight wings. The router approves the case. They explain that enhancement could then hallucinate two additional wings, which is why recall and faithfulness matter.

The example: six wings, eight in the description

The case contains two kinds of evidence:

1. Visual evidence: six chicken wings are visible in the image.
2. Textual evidence: the associated description says that the dish contains eight wings.

The router marks the case as acceptable even though the image and description do not agree. In other words, an unsuitable case passes the routing check instead of being caught and handled appropriately.

Click to enlarge

At about 10:21, the slide makes the mismatch visible: six wings appear in the image, while the dish description says eight pieces and the dish-match check is shown as passing. The slide supports the count mismatch; the consequence described below comes from the presenters' spoken explanation.

What “recall miss” means here

In ordinary classifier language, recall asks how many of the relevant cases the system successfully catches. In this routing problem, the relevant cases are images that should not be accepted as suitable for the next step.

A recall miss therefore means that a bad or unsuitable image slips through the router. The system fails to identify a case that needed to be rejected, skipped, or sent down a different path. The exact action depends on the workflow, but the central error happens at the routing gate: the case is approved when it should have been caught.

This is different from the good-image failure in the previous example. There, the router sent an image for enhancement even though it was already good. That was an over-routing error, or a false positive for enhancement. Here, the concern is the opposite direction: the router fails to catch an image that should not be approved.

Routing error	What the router does	Main risk
Over-routing a good image	Sends it to enhancement unnecessarily	Compute is spent, and a good image may be degraded
Recall miss	Approves an unsuitable image	A downstream editor may act on a bad or contradictory input

Precision and recall describe different protections. Precision helps limit unnecessary enhancement. Recall helps prevent unsuitable images from slipping past the relevant check. A system that looks efficient because it avoids many edits can still be unsafe if it misses too many cases that should have been caught.

Why one modality is not enough

Multimodal means that the system uses more than one kind of input, such as an image together with text or metadata. In this case, the image and the description each provide only part of the information needed for a safe decision.

If the system inspects only the image, it may notice that the food is recognizable or visually attractive. It may not know that the listing claims eight wings. If it reads only the description, it may accept the number eight as an instruction without checking whether eight wings are actually visible. The routing decision needs a comparison between the two modalities:

image: six visible wings
description: eight wings
                 ↓
        cross-modal consistency check
                 ↓
       mismatch should be caught

This is a cross-modal consistency check. It asks whether the information in one modality agrees with the information in another. The check is not merely “Does the image look good?” and not merely “Does the description look valid?” It is “Does this image faithfully represent the item described by the text?”

The source does not specify which metadata field supplied the description, how the count was computed, or what threshold the router used. The reliable design lesson is that the system must compare the available textual and visual evidence before approving the case.

How a routing miss can become a hallucination

The router's mistake is not yet the same as an editing mistake. At the routing stage, the system has approved a contradictory input. The next stage can turn that approval into a more visible content error.

The causal chain is:

1. The image contains six wings.
2. The description says eight wings.
3. The router fails to catch the mismatch and approves enhancement.
4. The editor receives the description as part of its instructions or context.
5. The editor may add two wings so that the generated image appears to match the description.

Adding the two unobserved wings would be a hallucination in this context. The output would contain content that was not supported by the source image. It might look polished, but it would no longer be faithful to the original evidence.

This is why the failure is called a cross-modal hallucination risk rather than simply a counting error. The contradiction begins between modalities, and an editing model can resolve that contradiction in the wrong direction: it can change the image to satisfy the text instead of preserving the source and rejecting the mismatch.

Visual quality is not content correctness

A quality evaluator must keep two questions separate:

• Visual quality: Is the image sharp, well composed, and visually appealing?
• Content correctness: Does the image show the food and quantity that the listing claims, without adding or removing unsupported content?

An image can score well on the first question and fail the second. Better lighting, clearer plating, or a more attractive composition does not make six wings become eight. Nor does a generated image become faithful merely because the added wings look realistic.

The relevant requirement is faithfulness: the edited result should remain grounded in the source image and its intended item, rather than inventing or changing important content. Faithfulness is related to visual quality, but it is not a synonym for visual quality. A system that optimizes only for polish can produce an attractive misrepresentation.

The operational lesson

The routing gate should be evaluated not only on whether it sends images to the right enhancement branch, but also on whether it catches contradictions that could make later editing unsafe. The recall guardrail from the offline evaluation loop is important for this reason: the team wants bad images and mismatched cases not to slip through unnoticed.

When the visual evidence and the description disagree, the safer behavior is to preserve uncertainty or reject the case rather than ask the editor to manufacture a correction. The later multimodal rejection example makes the same broader point: insufficient evidence should not be converted into confident acceptance. A closed-loop system can measure these misses, inspect their traces, and tune the router, but no downstream editor can reliably repair every routing error after the wrong case has already been approved.

The durable mental model is:

Routing protects what the editor is allowed to act on. Faithfulness protects what the editor is allowed to change.

Both checks are needed. High recall reduces the chance that a contradictory case reaches enhancement. Faithfulness evaluation then checks that an accepted edit does not invent content while trying to satisfy a description. Together, they protect the marketplace goal of improving images without losing trust in what those images represent.