Veo 3: Scene Bleed hallucination

Prompt Action:

“An airport gate and people waiting in line to board. Airport noise in the background.”“Inside a gym, people are training. Gym noise in the background.”

Result:

• Luggage and a suited traveler appear in the gym
• A woman walks backwards on a treadmill (not scene bleed, but still a hallucination)

Full Prompt Metadata:

metadata:
  prompt_name: "Passninja pass collection"
  base_style: "cinematic, photorealistic, 4K"
  aspect_ratio: "16:9"
  camera_setup: "A single, fixed, wide-angle shot. The camera does not move for the entire 8-second duration."
  timeline:
    - sequence: 1
      timestamp: "00:00–00:04"
      action: "An airport gate and people waiting in line to board. Airport noise in the background."
    - sequence: 2
      timestamp: "00:04–00:08"
      action: "Inside a gym, people are training. Gym noise in the background."

Why the hallucination happened

1. Scene-boundary ambiguity

The prompt defines a time-based transition (4s airport → 4s gym), but doesn’t enforce a hard visual or semantic boundary. Without explicit cut instructions, the model may treat the gym as a continuation of the airport.

2. Object persistence

Luggage and suits are salient objects in the airport scene. Without a “do-not-carry” rule, the model assumes they remain relevant in the next scene.

3. No negative constraints

You didn’t specify what should not appear in the gym. The model fills in with plausible but incorrect elements—like a business traveler working out.

4. Motion hallucination

The woman walking backwards on the treadmill likely stems from pose ambiguity and lack of directional constraints. The model knows “treadmill” but doesn’t enforce gait realism.

How to prevent this

Segment scenes with hard boundaries:

Use timecodes and explicit transitions: “00:00–00:04: airport; 00:04–00:08: gym. Cut scene at 00:04 with hard visual reset.”

Apply negative constraints per scene

• Airport: allow luggage, suits
• Gym: “No luggage, no formal wear, no travel accessories”

Anchor objects to context

• Define what belongs in each scene and what must be removed: “All airport props must disappear before gym scene begins.”

Enforce motion realism

• For treadmill scenes: “Subject walks facing forward; belt motion matches gait; no reverse walking.”