Introduction

You write a prompt, generate it, get a great clip. You write the same prompt again — and the result looks different. The character's hair changed, the lighting shifted, the camera angle slipped.

This is the most frustrating part of working with VEO 3, and it's not a bug — it's how the model works. This article explains why, and what to do about it.

Why Same Prompt ≠ Same Result

VEO 3 (and Veo 3.1) is a probabilistic model. Every generation samples from a distribution of possible outputs. Two things drive the variation:

• A random seed (unless you fix it)
• Tiny prompt parsing differences

Small ambiguities in your prompt let the model fill in the gaps differently each time. The fix is to remove ambiguity.

The Five Things to Lock Down

1. Character description

Write 4 lines covering build, clothing, hair, expression. Reuse word-for-word every time.

2. Lighting

Be specific: "golden hour with soft warm light from the left" beats "warm lighting".

3. Camera direction

Pick one angle, one shot type, one movement — and write all three.

4. Style / genre

Cinematic vs documentary vs anime makes a huge difference. Always specify.

5. Color palette

Even one color line stabilizes the output: "warm muted oranges and soft greens".

The Copy-Paste Solution

Build a "style block" you append to every prompt:

Style: cinematic film, soft warm lighting from the left, golden-hour color
palette with muted greens. Camera: low-angle tracking shot, slight
handheld feel. Color palette: amber, faded sage, muted cream.

Same block, every prompt in a series. Drift drops dramatically.

Use JSON Prompts

JSON prompts force structure and reduce parsing ambiguity. The full guide is at VEO 3 JSON prompting. Two prompts with identical JSON fields will give you nearly identical aesthetics.

Use Reference Images Where It Matters

For character consistency specifically, the most reliable trick is Frames-to-Video with a reference image:

• Generate one strong shot of your character
• Save the frame
• Upload it as a reference for subsequent clips

Catch: audio is disabled in Frames-to-Video mode. Use it for visuals only.

What VEO 4 Will Fix

The leaks suggest scene IDs / character anchors in VEO 4 — tag a character once, reuse them reliably across generations. This is the long-term solution. See Google VEO 4 leaked features.

Until then, the discipline above is your best friend.

Quick Checklist

• ✅ Lock your character description
• ✅ Lock lighting + style + camera + color palette
• ✅ Use a reusable "style block"
• ✅ Use JSON prompts
• ✅ Use reference images for tight character control

Conclusion

VEO 3 isn't broken when it gives you different results — it's working as designed. Lock your prompt structure, reuse your style block, and lean on JSON. Drift drops to almost zero.