Introduction

Once you've got the basics of writing prompts down, the next step is learning how to combine all of Omni Flash's input types into a single coherent prompt. Instead of relying entirely on text, you can upload a reference image for the character, a voice sample for the tone, a rough video for the camera move, and a short text prompt for the action.

In this article, I'll share the multimodal prompting framework I use for almost every Omni Flash generation. It works for cinematic shorts, social content, explainer videos, avatar-driven scenes, and basically anything you'd want to make.

If you haven't read the previous articles, start with the basics of writing Omni Flash prompts first.

The Framework

Here's the full structure:

[TEXT: character + action + camera + lighting + style + motion + environment + color]
[IMAGE: character reference OR style reference OR scene reference]
[VIDEO: existing clip to edit or extend]
[AUDIO: voice sample, music reference, or ambient sound cue]

You don't need to use all four input types every time. In fact, most of my best generations use just text + one image reference. The key is knowing which input type to lean on for which creative element.

1. Text: The Director's Instructions

Text is still the backbone of your prompt. It tells the model what to do with all the other inputs. Even when you upload references, the text prompt controls:

• Action — what the character is doing
• Camera — angle, motion, distance
• Lighting — time of day, light source, mood
• Motion — environmental movement, physics
• Audio direction — ambient sound, music, dialogue tone

The rule is simple: use text for things that change, and images for things that stay the same.

For example, if your character's face should stay identical across generations, upload a reference photo and keep the text prompt focused on what they're doing and where the camera is.

Example text prompt:

A young woman stands at the edge of a cliff at sunset. The camera pushes
in slowly from a low angle behind her. Warm golden light over the ocean
below. Her white dress billows in the wind. Style: indie drama film.
Color palette: warm yellows, faded greens, soft peach tones. Wind howls
in the background with soft ambient piano music.

Notice what's NOT in the text: a detailed description of her face, hair, or exact dress design. If we want those locked, we'll upload a reference image instead of describing them.

2. Image References: Lock the Visuals

Images are the fastest way to get exact visual control. Omni Flash treats uploaded images as unified scene descriptions rather than separate ingredients.

Character Reference

Upload a photo or illustration of the exact face, body type, or outfit you want. The model will match it more closely than any text description.

When to use:

• You need a specific person (your avatar, a known actor lookalike)
• You want consistent facial features across multiple generations
• You have a sketch or concept art that nails the look

Best practices:

• Use clear, well-lit frontal photos for faces
• Include the full outfit in the frame if clothing matters
• Avoid cluttered backgrounds in reference images

Style Reference

Upload a screenshot from a film, a painting, or another AI-generated image that has the exact color grading, texture, or mood you want.

When to use:

• You want a specific film's look (e.g., "Blade Runner 2049" lighting)
• You're matching existing brand visuals
• You need a specific artistic medium (claymation, watercolor, etc.)

Scene Reference

Upload a photo of a real location or a 3D render of the environment.

When to use:

• You're recreating a real place (your office, a specific city street)
• You want architectural accuracy
• You have a mood board photo that captures the exact environment

⚠️ Don't upload more than 2–3 image references in one prompt. The model's ability to weight multiple visual inputs decreases as you add more. Pick your strongest 1–2 references and let text handle the rest.

3. Video References: Edit What Exists

This is one of Omni Flash's standout features. You can upload an existing video clip and ask the model to edit, extend, or restyle it.

Use cases:

• Cinematic remake — upload a phone video and ask for "cinematic color grading, shallow depth of field, 24fps film look"
• Style transfer — upload a clip and ask for "convert to 2D anime style while keeping the same camera moves"
• Extension — upload a 5-second clip and ask for "continue this scene for 5 more seconds with the same character"
• Object swap — upload a clip and ask for "replace the red car with a blue motorcycle, keep everything else"

Best practices:

• Keep the uploaded clip short (under 5 seconds works best)
• Be specific about what to change AND what to preserve
• Use conversational editing for multi-step changes

Example prompt with video reference:

Take this clip and restyle it as a gritty noir film. Add harsh
black-and-white contrast, deep shadows, and cigarette smoke in the
background. Keep the same camera angle and character position.

4. Audio References: Set the Mood

Omni Flash can take audio inputs to guide the sound design of the generated video. This is less about uploading a full soundtrack and more about giving the model a reference for tone, rhythm, or voice quality.

Use cases:

• Voice reference — upload a short speech sample so the generated dialogue matches a specific vocal tone or accent
• Music reference — upload a 5-second music clip to set the genre, tempo, and mood of the background score
• Ambient reference — upload rain, city traffic, or forest sounds to guide the environmental audio

Best practices:

• Keep audio references under 10 seconds
• Use clean, isolated samples (no background noise competing with the main sound)
• Describe in text what you want the audio to do: "match the rhythm of this music clip" or "use this voice tone for the dialogue"

Example prompt with audio reference:

A drummer performs on a rooftop at dusk. The camera orbits slowly
around them. Style: cinematic music video. Match the rhythm and energy
of this audio clip for the drumming and background music. Warm sunset
lighting, city skyline behind them.

Putting It All Together: Full Examples

Example 1: Text + Image (Character Reference)

Image uploaded: Clear photo of a young man with short black hair, wearing a leather jacket.

Text prompt:

He walks slowly through a rainy Tokyo alley at night. Neon signs
reflect in puddles around his feet. The camera follows behind him
in a low tracking shot. Style: moody cinematic thriller. Soft synthwave
music plays in the background. Color palette: neon pink, deep blue,
wet asphalt gray.

What the image handles: exact face, hair, outfit. What the text handles: action, camera, environment, style, audio, color.

Breakdown:

• Character: reference image (young man, leather jacket)
• Action: walks slowly through a rainy alley
• Camera: low tracking shot from behind
• Environment: Tokyo alley at night, neon signs, puddles
• Style: moody cinematic thriller
• Audio: soft synthwave music
• Color: neon pink, deep blue, wet asphalt gray

Example 2: Text + Video (Cinematic Remake)

Video uploaded: 4-second phone clip of someone pouring coffee in a kitchen.

Text prompt:

Restyle this clip as a cinematic food documentary. Add warm overhead
lighting, shallow depth of field, and steam rising from the cup. The
camera should slowly push in on the pour. Sounds of a quiet morning
kitchen: soft clinking, distant birds outside. Color palette: warm
browns, cream, soft gold.

What the video handles: base action, original camera position, real-world physics. What the text handles: lighting, depth of field, camera motion, sound design, color, style.

Example 3: Text + Image + Audio (Full Multimodal)

Image uploaded: Photo of a middle-aged woman in desert robes, sun-weathered face.

Audio uploaded: 5-second clip of ambient desert wind with a distant string instrument.

Text prompt:

She stands alone on a sand dune at dawn. The camera starts as a wide
shot, then slowly pushes in to a medium close-up. The rising sun casts
long shadows across the dunes. Style: epic historical film. She looks
out at the horizon with a determined expression. Match the mood and
instrumentation of this audio clip for the background score. Color
palette: warm golds, burnt orange, deep sand brown.

What the image handles: exact character look. What the audio handles: mood, instrumentation, musical tone. What the text handles: action, camera, environment, style, color, facial expression.

Example 4: Conversational Editing Loop

One of Omni Flash's most powerful workflows is the multi-turn edit. Here's how a full session might look:

Turn 1 — Initial generation:

A young astronaut floats weightlessly inside a space station module.
The camera is a slow orbit around them. Soft blue interior lighting
from control panels. Style: realistic sci-fi. Ambient hum of machinery
in the background.

Turn 2 — Lighting adjustment:

Make the lighting warmer — add a golden sunset glow coming through
a viewport. Keep the same camera orbit and character position.

Turn 3 — Prop addition:

Add a floating holographic Earth globe spinning slowly near their
hand. Same lighting, same camera, same character.

Turn 4 — Audio layer:

Add soft orchestral music swelling slowly underneath the ambient hum.
Keep everything else the same.

Each turn preserves the scene state and builds on it. You end up with a complex, layered shot without ever rewriting the full prompt.

Common Multimodal Mistakes

Here are the pitfalls I see most often:

• Overloading with references — 4 images + a video + an audio clip + a novel-length text prompt overwhelms the model. Start with 1–2 inputs.
• Conflicting directions — the image reference shows bright daylight, but the text says "night scene." The model will struggle. Align your inputs.
• Vague image uploads — blurry, dark, or cluttered reference images hurt more than they help. Use clean, well-lit references.
• Expecting exact audio replication — the model uses your audio reference for mood and tone, not as a sample to copy precisely.
• Forgetting to describe audio in text — even with an audio reference, add a text line about what you want: "match the rhythm of this clip" or "use this voice tone."
• Editing without state preservation — in conversational editing, always remind the model what to keep: "same character, same room, but change the lighting."

A Practice Routine

Here's how to build multimodal prompting skill:

1. Start with text only. Write the best text prompt you can for a single scene.
2. Add one image reference. Upload a character photo. Compare the output to your text-only version. Notice what improved.
3. Swap the image. Use a different character or style reference. See how much the scene changes while the action stays the same.
4. Try a video reference. Upload a rough clip and ask for a restyle. Notice how the model preserves motion while changing aesthetics.
5. Add audio. Upload a short music clip and compare the sound design to a text-only audio description.

Do this five times with different subjects, and you'll have an intuitive sense of which input type to reach for in any situation.

Conclusion

Multimodal prompting isn't about using every input type at once. It's about choosing the right tool for each creative element and letting text handle the direction while references handle the visuals and sound.

Text for action, camera, and mood. Images for faces, styles, and places. Video for edits and extensions. Audio for tone and rhythm. That's the whole framework.

If you want to start experimenting, you can plug your multimodal prompts directly into our Veogen Studio Gemini Omni Flash generator. Upload your references, write your direction, and iterate.