Multimodal model overlooks clearly provided frames, misses visible duplicate characters, and hallucinates events when analyzing image sequences

I encountered recurring problems when asking DeepSeek to analyze sequences of images from my fictional PowerPoint videos. The main issue is not simply that the model occasionally misunderstands an image. It sometimes fails to account for frames that were clearly provided, overlooks characters or actions visibly present in those frames, and then fills the resulting apparent gaps with plausible narrative events.

1. Failing to recognize visible actions and then incorrectly claiming the relevant frames were absent

In one video, “4 Ways to Bit Leo,” I provided all of the frames and explicitly stated that the frames were incorrectly ordered.

The fourth sequence, “Cockroach Chocolate,” visibly depicted:

Leo accidentally eating chocolate containing a cockroach;
green liquid coming from Leo's mouth;
green liquid remaining on the floor;
Leo slipping and falling afterward.

DeepSeek omitted the vomiting and slipping from its reconstruction. When I pointed out the omission, it responded:

“I actually don't see the vomiting frame in that batch of screenshots—it must have been cut out or is part of a slide you didn't include.”

After I explained the sequence, it again said:

“I was so focused on the specific frames you sent (which didn't include the green liquid or the slipping)…”

However, those frames were included from the beginning.

The problem was therefore not merely that the model called vomit a “green oval.” It failed to recognize an action that was visibly depicted and then treated its own failure to recognize the evidence as evidence that the relevant frames had not been provided.

If the meaning of the green liquid was genuinely unclear, a better response would have been:

“I can see green liquid coming from Leo's mouth and later on the floor, but I am not certain what it represents. Is this supposed to be vomiting?”

The model should not claim that a frame or action was absent merely because it failed to interpret it.

2. Missing duplicate characters that are clearly visible simultaneously

Some frames contained two duplicate characters wearing identical black clothes at the same time.

DeepSeek initially failed to acknowledge that both black-clothed characters were simultaneously present. It later explained the problem by saying:

“Since you used multiple cutout versions of the same person (Black Aayan 1 and 2), I couldn't keep track of who was who.”

This does not fully explain the error.

Tracking identical characters across different frames may genuinely be difficult, but when two separate copies are clearly visible simultaneously, the model should first recognize that there are two visible instances.

For example, it could say:

“I can see two visually identical black-clothed characters simultaneously, one on the left and one on the right. I cannot yet determine which one should be labeled Black Aayan 1 or Black Aayan 2 across the full sequence.”

The model should distinguish between:

detecting that two characters are present in the current frame, and
tracking which duplicate is which across multiple frames.

The first task should not fail simply because the characters look identical.

3. Inventing actions that were not shown in the frames

In another video, “Aayan eats angela birthday cake,” Black Aayan 2:

Cut a slice of cake.
Left the slice on the table and left the scene.
Came back.
Took the slice he had left.
Took the entire cake.

The frames did not show Black Aayan 2 eating the slice.

Nevertheless, DeepSeek repeatedly inserted eating into the sequence. During one supposed review of the images, it wrote:

“Image 18: Black Aayan 2 is on the right, eating.”

This action was not visibly depicted.

The model appears to have transformed a plausible expectation—

“Someone who takes a slice of cake will probably eat it”

—into a claimed visual fact.

If the model cannot determine what happens, it should say so or ask for clarification. For example:

“I can see Black Aayan 2 take the slice, but I do not see an eating action. Does he eat it, leave with it, or do something else?”

The model may consider expected narrative events as possibilities, but it should not present them as facts unless they are actually visible or explicitly established by the user.

4. Treating a provided image sequence as though some frames do not exist

In older parts of the conversation, DeepSeek sometimes stated that a specific frame or event was not present even though the relevant frame had clearly been sent.

This is particularly frustrating because the user has already supplied the evidence. Instead of saying:

“I may have overlooked or misunderstood that frame,”

the model may say:

“That frame was not included.”

This can create a chain of errors:

Frame is provided → model fails to recognize or integrate it → model believes there is a gap → model either claims the frame is missing or fills the gap with a plausible event.

The model should reliably account for all provided frames. If it cannot determine how a frame fits into the sequence, it should explicitly state that uncertainty rather than treating the frame as absent.

5. Failure to reconstruct chronology despite explicit plot constraints

In “Aayan destroys electronics - Tablet!”, I explicitly provided the chronological plot:

Aayan drops the tablet onto the ground.
He throws the tablet backward using his leg.
He rides a bicycle into the tablet.
He smashes the battery and the rest of the tablet.

Despite having this chronological information, DeepSeek initially produced a sequence that placed the hammer/battery destruction before the earlier drop, leg throw, and bicycle actions.

It appeared to describe or group frames without first reconstructing their chronological relationship.

For image-sequence analysis, the model should first:

Inspect all available frames.
Identify the visible state of important objects and characters in each frame.
Compare changes between frames.
Group frames into events.
Use explicitly supplied plot information as chronological constraints.
Only then produce a chronological narrative.

Recognizing individual objects is not sufficient if the model fails to integrate them into the correct timeline.

6. Excessive confidence after uncertain or incorrect reconstruction

DeepSeek repeatedly described its answers as:

“100% accurate”

“100% correct”

“I finally have it perfectly right”

even after previous supposedly final reconstructions had contained major errors.

This makes the problem worse because the model is not only making an incorrect interpretation; it is expressing certainty that is not supported by its visual analysis.

When ambiguity remains, the model should distinguish between:

directly visible facts;
inferences;
uncertain interpretations;
information that cannot be determined from the frames.

It should avoid declaring a reconstruction completely accurate unless there is sufficient evidence.

Expected behavior

For image-sequence analysis, I would expect the model to:

Account for every provided frame rather than treating overlooked or misunderstood frames as absent.
Recognize multiple visible duplicate characters even when their identities are difficult to track across frames.
Describe directly visible actions separately from inferred events.
Never turn a plausible narrative expectation into a claimed visual fact without evidence.
Ask for clarification when an important action or symbol cannot be determined.
Reconstruct chronology by comparing all frames and using any plot information supplied by the user.
Avoid claiming that an image or event was missing unless it is actually certain that it was not provided.
Calibrate confidence and avoid “100% accurate” claims when the reconstruction contains unresolved ambiguity.
Main concern

The model should not respond to uncertainty by silently inventing a more familiar story.

A robust approach would be:

Visible evidence → state it as observation.
Possible interpretation → label it as an inference.
Unclear or missing understanding → say so and ask for clarification if necessary.

The most problematic pattern I encountered was:

Clearly provided frame → model fails to recognize or integrate it → model treats it as absent → model fills the apparent gap using an expected narrative event → model presents the invented reconstruction with high confidence.

This is especially problematic for fictional animations whose actions may deliberately differ from ordinary real-world or conventional narrative expectations. The model should follow the visual evidence and the user's established plot rather than replacing unusual fictional events with statistically more familiar ones.