Veo 3: Same Prompt Totally Different Videos – Hallucination

This example clearly illustrates a common AI hallucination pattern: same input, divergent outputs.

Heads-up: this is just a reference example. For the actual video, I developed a comprehensive plan and pre-defined many elements in advance. So while this prompt isn’t central to the final output, I’m using it here because it illustrates an important point: over-defining items without anchoring them to specific actions or constraints can lead to hallucinations.

Video Breakdown:

First 8 seconds: One video interpretation of the prompt.

Next 8 seconds: A completely different video generated from the same prompt.

Final 8 seconds: Both videos side by side—revealing stark differences in character, setting, and motion.

Here is the prompt I used(I went full JSON for this one—not recommended).

[
  {
    "character": {
      "name": "Leo",
      "age": 26,
      "gender": "male",
      "appearance": {
        "build": "lean and athletic",
        "features": {
          "hair": "dark, tousled, sometimes tied with a bandana",
          "facial_hair": "minimalist beard or clean-shaven",
          "eyes": "focused, alert gaze",
          "skin": "tanned with subtle sun marks"
        },
        "attire": {
          "top": "moisture-wicking paddle shirt",
          "bottom": "high-performance shorts",
          "footwear": "lightweight sneakers with traction soles",
          "accessories": [
            "performance watch",
            "sweatband",
            "wrap-around sunglasses"
          ]
        },
        "gear": {
          "racket": "sleek carbon-fiber paddle racket",
          "bag": "minimalist gym bag with NFC pass sleeve"
        }
      },
      "energy": {
        "movement": "explosive footwork, quick directional shifts, fluid transitions",
        "warmup": [
          "lunges",
          "paddle-specific stretches",
          "short sprints"
        ]
      },
      "personality": {
        "vibe": "competitive yet laid-back, philosophical, high-spirited",
        "music_preference": "Latin pop and high-tempo EDM",
        "sportsmanship": "fist-bumps opponent, embraces challenge"
      },
      "cinematic_highlights": [
        "slow-motion sweat droplets mid-sprint",
        "shadow stretching across the court at sunset",
        "dust kicked up during sidestep",
        "zoom-in on intense pre-serve gaze"
      ]
    }
  },
  {
    "sequence": "arrival",
    "location": "paddle facility entrance",
    "time_of_day": "late morning",
    "actions": [
      "Leo walks toward the paddle office with relaxed athletic confidence",
      "Swipes NFC pass at entry panel, door slides open",
      "Brief nod at the front desk, no dialogue",
      "Adjusts sweatband and checks paddle racket grip"
    ],
    "camera_style": {
      "movement": "smooth pan followed by tight tracking shot",
      "focus": [
        "sunglasses reflection",
        "NFC swipe",
        "gear reveal"
      ],
      "lighting": "natural light with sharp indoor transition"
    },
    "ambient_details": {
      "background_sound": [
        "distant paddle hits",
        "ambient music"
      ],
      "visuals": [
        "digital match listing showing Leo's name"
      ]
    },
    "character_mood": "calm, focused, in sync with environment"
  }
]

What went wrong: Accessories and gear not associated to a character or action

Defined, but not fixe: I listed accessories like a performance watch, sweatband, wrap-around sunglasses, and gear like a paddle racket and gym bag. But without:

• Color
• Placement (e.g. “watch on left wrist”)
• Visibility rules (e.g. “sunglasses off indoors”)
• Scene-specific usage (e.g. “grips racket before entering court”)

…the model treats them as optional visual flourishes rather than essential identity markers.

No action linkage Accessories weren’t tied to specific behaviors. For example:

• “Leo adjusts his sweatband” is good — it anchors the item to a moment.
• But if the sunglasses or bag aren’t mentioned in any action, the model may omit them, change them, or invent substitutes.

No continuity enforcement Without constraints like “accessories must appear in all scenes unless removed on-camera,” the model assumes it can vary them freely between shots.

How to fix it

To prevent this kind of drift:

1. Lock accessory attributes

• Specify color, material, and placement: “Black sweatband on right wrist; silver watch on left wrist; sunglasses worn only outdoors.”

2. Tie gear to actions

• Make each item part of a behavior: “Leo adjusts his sweatband before entering; grips paddle racket with right hand; slings gym bag over left shoulder.”

3. Use continuity rules

• Declare what must persist: “All accessories must remain consistent across scenes unless removed on-camera.”

4. Add negative constraints

• Prevent substitutions: “No hats, no tennis rackets, no duffel bags.”

The takeaway

You’re absolutely right: when accessories and gear are defined but not grounded, they become sources of hallucination. The fix isn’t just more detail — it’s anchoring those details to actions, positions, and continuity rules so the prodigy knows what to preserve.

If you’d like, I can help you rewrite your original prompt with these constraints built in — or create a reusable “character continuity” template for future video work. Want to build that next?