In one video, a teacher raised a red notebook near her chin. In another, she held it diagonally across her chest. In a third, she shifted both hands before lifting it. I had submitted the same reference image and the same prompt three times, without changing the visible settings. Why did the results differ?
This matters when a scene must do more than look plausible. A teacher who turns her head is not necessarily looking at the whiteboard she was asked to face. A notebook that stays red may still be held differently. To diagnose the difference, I needed to separate what stayed consistent from what the prompt did not control precisely.
The reference image shows a teacher in a blue cardigan holding a red notebook in a classroom. I used the same prompt for all three runs. Its visual and motion instructions were:
A female teacher in a blue cardigan holds a red notebook in a classroom. She gently raises the red notebook and turns her head toward the whiteboard. Static camera, natural and subtle classroom motion. Preserve her face, clothing, the red notebook, the whiteboard, and the classroom background from the reference image. No dialogue.
Figure 1. The same reference image was used without replacement or re-uploading.
First, check whether the inputs really stayed the same
I made three independent Image to Video submissions through Vidu Q3 in one session. The interface showed Vidu Q3 Pro Fast, 720P, 5 seconds, and Movement set to Auto Move. Audio FX and BGM were off. I left the image in place, did not revise the prompt, and downloaded each completed video before starting the next run.
The account began with 40 credits. Each submission cost 10, leaving 10 after all three. Every generation completed and downloaded; there was no failed run or retry. That gives me three files produced under the same visible conditions, not proof that every hidden generation variable was identical. I saw no seed control in this mode, so I cannot claim the runs shared a seed. The mode also showed no separate aspect-ratio selector; the downloaded files were all 16:9.
Figure 2. The recorded inputs and visible settings before the first run. The full prompt is reproduced above.
Next, compare the action at matching times
The three files matched on basic specifications: each was approximately 5.04 seconds long, 1280 × 720, and 24 frames per second. None had a detected audio track, which is consistent with the sound options being off. The checksums differed, but a different checksum alone does not explain which visual requirements changed.
I extracted frames every 0.5 seconds from 0 through 4.5 seconds and added a frame near 4.95 seconds. Looking at matching time points made the differences easier to describe:
Run 1: The teacher turned toward the left side of the frame, then raised the notebook near her chin at about 3–3.5 seconds. Her profile looked closer toward the end.
Run 2: Her body turned farther. At about 3.5 seconds, she held the notebook diagonally across her chest, with a different grip and crossed-arm position.
Run 3: She adjusted both hands around 2.5–3 seconds, supported the notebook near its bottom, and lifted it around 3.5–4 seconds. The final framing was wider, revealing more of her hands and waist.
Figure 3. Frames at 0, 2.5, and approximately 4.95 seconds. The character and notebook remain recognizable, but the action and composition do not repeat exactly.
All three clips retained the broad task: the teacher remained recognizable, the cardigan was blue, the notebook stayed red, and she turned her head and lifted the notebook. The diagnostic point is narrower. The timing, hand placement, body angle, and framing were not interchangeable. If an edit needs a particular grip or movement at a particular moment, “the action happened” is not a sufficient pass condition.
Do not confuse a head turn with the requested direction
The prompt specified a turn toward the whiteboard. In the reference image, the whiteboard was behind the teacher and toward the right side of the frame. In all three outputs, she looked toward the left side of the frame. The video showed a head turn, but I could not count it as a clear turn toward the board.
That distinction changes the next step. Generating more versions under the same instructions might yield a more attractive movement, but this small test gives no basis to promise that it will correct the spatial direction. In a later test, I would clarify the intended screen direction and assess that requirement on its own.
The background needed a separate check. In Run 2, the poster that had four geometric shapes in the reference gained a purple circle and a yellow square, bringing the visible total to six. That matters if the poster carries lesson information, even if the teacher and notebook look acceptable.
Figure 4. Near 4.95 seconds in Run 2, the poster includes two shapes absent from the reference image.
Check the settings before blaming the prompt—or the model
I asked for a “Static camera,” but the whiteboard, shelves, desks, and poster shifted in the frame across the videos. It would be tempting to label that a straightforward failure of the written instruction. My test has a complication: I also left the interface’s Movement control on Auto Move. The camera instruction and movement setting may have pulled in different directions. These runs show that the framing was not static; they do not isolate the cause.
The same caution applies to repeatability. Three different results under the same visible inputs show variation in this setup. Because I did not control or observe a seed, I cannot identify the internal reason for each difference or calculate how often it would happen in other sessions, modes, or services.
For a useful diagnosis, I would ask five questions at the same time points in each clip: Is the face and clothing recognizable? Does the notebook keep its shape and color? When does the lift begin and how is it gripped? Does the gaze actually reach the whiteboard? Do the poster and furniture stay where they should? A requirement should be marked as met only when the relevant frames support it.
If I ran a follow-up experiment, I would save the same image and prompt, record every visible setting, and change just one variable at a time. The first candidate would be to make the camera instruction and Movement setting agree. I would then check whether that changed the framing, without calling it a proven fix in advance.
These three samples are enough to reject the assumption that identical visible inputs gave me identical clips. They are not enough to establish a general success rate or prove an internal mechanism. My review used half-second key frames plus a final frame, not a frame-by-frame certification or a separate quality review of continuous playback. It was also a test through an independent web service, not an official test by the model developer.
The practical lesson is to diagnose the specific requirement that moved: identity, object, timing, direction, camera, or background. In this case, the character and notebook largely survived; the exact movement and spatial details did not. If you repeated a short AI scene, which difference would make you reject the result first?