Models / Veo / Reference to video

Veo reference to video prompting guide

How to prompt Veo in reference-to-video mode. Every point links its source; community tips are marked. Your agent gets the same guide through the Atlas MCP (get_guide).

Prompt structure

cinematography→subject from refs→action→context→style ambiance→audio

• Five-part formula: [Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance]; cinematography is the most powerful element for tone. source ↗

What the text should describe

• referenceImages: up to three images as style and content references, each with referenceType 'asset'; requires durationSeconds 8; not on Veo 3.1 Lite. source ↗
• 'Ingredients to video': reference images of characters, objects, settings or style keep them consistent across shots; Google pairs it with Gemini 2.5 Flash Image to make the references first. source ↗

What not to describe

• Don't expect a tagging syntax: Google's example prompt describes the referenced dress, sunglasses and woman in prose, with no @tags. source ↗

Camera vocabulary

• Movement: dolly, tracking, crane, aerial, slow pan, POV. Composition: wide, close-up, extreme close-up, low angle, two-shot. Lens: shallow depth of field, wide-angle, soft focus, macro, deep focus. source ↗

Responds well to

• Audio: dialogue in quotes (A woman says, "..."), 'SFX: ...', 'Ambient noise: ...'. source ↗

Duration and limits

• Text input limit 1,024 tokens. durationSeconds 4, 6 or 8; must be 8 with reference images, extension, or 1080p/4k. Aspect 16:9 or 9:16. Extension: +7 s up to 20 times, up to 148 s total, 720p only, Veo-generated input only. source ↗

Multi-shot

• Timestamp prompting: [00:00-00:02] shot one ... [00:02-00:04] shot two, each with camera, action, SFX or emotion. source ↗

Avoid and failure modes

• Phrase exclusions positively: 'a desolate landscape with no buildings or roads' rather than 'no man-made structures'. source ↗
• Creators report an unwanted slow-motion feel on Veo 3.1; state real-time speed explicitly (community report, unverified). community source ↗

Negative prompt

Supported via negativePrompt. source ↗

Real prompts, credited

Prompt describes the referenced woman, dress and sunglasses in prose; referenceImages = [dress, sunglasses, woman], each referenceType 'asset'.

Next

Sources

• ai.google.dev/gemini-api/docs/veo
• cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1
• x.com/IamNiccoloRizzo/status/2100524532658544932