Gemini Omni Flash:
The Any-Input-to-Video Model

import base64
from google import genai

client = genai.Client()

# Step 1: Generate the initial video
interaction = client.interactions.create(
    model="gemini-omni-flash-preview",
    input="A marble rolling fast on a chain reaction track, continuous smooth shot."
)

# Step 2: Save the output
if interaction.output_video:
    with open("marble_v1.mp4", "wb") as f:
        f.write(base64.b64decode(interaction.output_video.data))
    print(f"Generated {len(interaction.output_video.data)} bytes")

# Step 3: Edit the video through conversation
edited = client.interactions.continue_interaction(
    interaction_id=interaction.id,
    input="Change the marble to a glowing neon sphere and add dramatic backlighting"
)

# Step 4: Apply a second edit
final = client.interactions.continue_interaction(
    interaction_id=interaction.id,
    input="Slow the motion to half speed and add a deep bass sound effect on impact"
)

# Save the final version (3 edits total: create + 2 refinements)
if final.output_video:
    with open("marble_final.mp4", "wb") as f:
        f.write(base64.b64decode(final.output_video.data))

# Install the Google Gen AI SDK
pip install google-genai

# Set your API key (get one free at ai.google.dev)
export GOOGLE_API_KEY="your-key-here"

# Verify the installation
python -c "from google import genai; print('SDK ready')"

# Upload a reference image for animation
uploaded = client.files.upload("reference_frame.png")

# Use the uploaded file as input to Omni
interaction = client.interactions.create(
    model="gemini-omni-flash-preview",
    input="Animate this scene: gentle wind through the trees, birds crossing the sky",
    files=[uploaded]
)

A white ceramic mug sits centered on a matte oak table. Camera slowly orbits 180
degrees around it. Soft studio lighting from the upper left. The mug has a clean
sans-serif logo. Shallow depth of field. No text overlays. Photorealistic.

Aerial drone shot pulling back from a solitary lighthouse on a rocky cliff at
golden hour. Atlantic ocean waves crash against the rocks below. Seagulls cross
the frame. Cinematic color grading with warm highlights and deep shadows. 16:9.

A modern mobile app login screen with a dark theme. The email and password fields
fill in with a typing animation. The login button pulses blue, then the screen
transitions to a dashboard with a smooth slide-up animation. 9:16 portrait.

Animate the attached product image: the item gently rotates 45 degrees on its
axis while soft particles of light drift upward in the background. Maintain the
exact product appearance and proportions. Loop-ready ending.

Animate the attached portrait: subtle breathing motion, a slight head turn to
the right, natural eye blink. Hair moves gently as if in a light breeze. Maintain
photorealistic quality and the exact facial features from the source image.

Turn 1: "A person walking through a busy Tokyo street at night, neon signs reflected in puddles."
Turn 2: "Make it look like a Blade Runner scene: heavy rain, more neon, volumetric fog."
Turn 3: "Add a holographic billboard in the background showing a Japanese character."

Turn 1: "A glass of red wine on a marble counter, overhead fluorescent lighting."
Turn 2: "Change the lighting to candlelight. Warm tones, flickering shadows."
Turn 3: "Add a second glass of wine being poured from a bottle on the right side."

Turn 1: "A golden retriever running through a field of sunflowers."
Turn 2: "Replace the dog with a fox. Keep the same running motion and field."
Turn 3: "Change the sunflowers to lavender and shift the time to early morning mist."

A heavy iron ball drops from 3 meters onto a wooden floor. Show the realistic
impact: the floor dents, dust rises, and the ball bounces twice with decreasing
height. Real-world gravity. Side camera angle. Slow motion at 120fps feel.

A glass bottle tips over on a white surface. Water pours out in a smooth stream,
pools on the surface, and spreads naturally. Show realistic fluid dynamics:
surface tension, reflection, and refraction in the pooling water. Top-down camera.

A Rube Goldberg machine: a marble rolls down a track, hits a lever that launches
a small ball into a cup that tips, pouring water onto a wheel that spins and
rings a bell. Continuous single shot. Physically accurate chain reaction.

# Agentic pattern: generate 5 product videos from a CSV of product images
for product in products:
    interaction = client.interactions.create(
        model="gemini-omni-flash-preview",
        input=f"Product showcase: {product['name']}. Slow 360-degree rotation on a "
              f"clean white background. Studio lighting. {product['style']} aesthetic.",
        files=[client.files.upload(product['image_path'])]
    )

# Generate a 9:16 Instagram Reel from a product image + brand audio
interaction = client.interactions.create(
    model="gemini-omni-flash-preview",
    input="Create a vertical social media reel: zoom into the product, pan across "
          "the texture, pull back to show the full item. Trendy transitions. "
          "Synchronized to the attached audio beat.",
    files=[product_image, brand_audio]
)

# Agentic pattern: generate, QA with vision model, auto-fix with edit
interaction = client.interactions.create(
    model="gemini-omni-flash-preview",
    input="A person typing on a laptop in a modern office, natural lighting."
)

# Use Gemini 3.6 Flash to QA the output
qa_result = client.models.generate_content(
    model="gemini-3.6-flash",
    contents=["Check this video for artifacts, inconsistent lighting, or "
              "unnatural motion. List any issues.", interaction.output_video]
)

# Auto-fix any issues found
if "lighting inconsistency" in qa_result.text:
    client.interactions.continue_interaction(
        interaction_id=interaction.id,
        input="Fix the lighting inconsistency: ensure uniform natural light from the window."
    )

# Full agentic pipeline architecture
# Agent 1 (Orchestrator): Decomposes the video production goal
# Agent 2 (Asset Gatherer): Generates reference images via Gemini
# Agent 3 (Video Producer): Calls Omni Flash to generate clips
# Agent 4 (QA Inspector): Runs vision-based quality checks
# Agent 5 (Editor): Feeds conversational edits back to Omni

class OmniVideoPipeline:
    def __init__(self):
        self.client = genai.Client()

    def produce(self, brief: str, reference_images: list[str]):
        # Generate initial video from the brief
        files = [self.client.files.upload(img) for img in reference_images]
        interaction = self.client.interactions.create(
            model="gemini-omni-flash-preview",
            input=brief,
            files=files
        )

        # QA loop: check and fix up to 3 times
        for edit_round in range(3):
            issues = self.qa_check(interaction.output_video)
            if not issues:
                break
            interaction = self.client.interactions.continue_interaction(
                interaction_id=interaction.id,
                input=f"Fix: {issues}"
            )

        return interaction.output_video