AI Video Generation

Quick Reference

Model Selection Logic

def analyze_prompt_complexity(prompt: str, topic: str = "") -> Dict:
    combined = f"{prompt.lower()} {topic.lower()}"

    # HIGH complexity → Sora 2 Pro
    high_keywords = ["transformation", "morphing", "cinematic", "epic", "dramatic"]
    for kw in high_keywords:
        if kw in combined:
            return {"complexity": "high", "model": "sora-2-pro", "duration": 8}

    # MEDIUM complexity → Sora 2
    medium_keywords = ["tracking", "dolly", "movement", "animation", "transition"]
    for kw in medium_keywords:
        if kw in combined:
            return {"complexity": "medium", "model": "sora-2", "duration": 8}

    # LOW complexity → Veo 3
    return {"complexity": "low", "model": "veo3", "duration": 6}

Smart Generation with Fallback

1. Analyze prompt complexity
2. Select model (Sora or Veo)
3. If Sora fails → Fallback to Veo
4. Return video path

Video Specifications

Instagram Reels Requirements

Veo Parameters

config = {
    "aspect_ratio": "9:16",      # or "16:9"
    "duration_seconds": 8,        # 4, 6, or 8
    "number_of_videos": 1,
    "person_generation": "allow_all"
}

Sora Parameters

params = {
    "model": "sora-2",           # or "sora-2-pro"
    "duration": 8,                # 4, 8, or 12
    "size": "720x1280"           # or "1280x720"
}

Prompt Formula

5-Part Formula:

[CINEMATOGRAPHY] + [SUBJECT] + [ACTION] + [CONTEXT] + [STYLE]

Example:

"Slow dolly shot, IoT sensor device on greenhouse shelf, LED indicator
blinking green, morning sunlight through glass panels, professional
documentary style, shallow depth of field"

Generation Flow

Veo Generation

from app.veo_helper import generate_video_veo3

result = await generate_video_veo3(
    prompt="Your detailed prompt here",
    aspect_ratio="9:16",
    duration_seconds=8
)

if result.get("success"):
    video_path = result["video_path"]

Smart Generation (Sora + Veo Fallback)

from app.sora_helper import generate_video_smart

result = await generate_video_smart(
    prompt="Your prompt",
    topic="Technology",
    duration=8
)

# Automatically falls back to Veo if Sora fails
video_path = result.get("video_path")
model_used = result.get("model_used")  # "sora-2" or "veo-3"

Environment Variables

GEMINI_API_KEY=your_google_api_key    # For Veo
OPENAI_API_KEY=your_openai_api_key    # For Sora

Timeouts and Limits

Kling AI (fal.ai)

Endpoints

Features

• Duration: 5s or 10s (longer than Veo/Sora!)
• Aspect Ratio: 9:16, 16:9, 1:1
• Tiers: Standard (fast), Pro (quality), Master (premium)

Kling Prompt Format

Formula:

Subject + Description + Movement + Scene + Camera + Lighting + Atmosphere

Example:

Medium shot, bokeh background, a technician in safety helmet, checking
sensor readings on tablet, industrial factory floor, warm ambient lighting,
professional documentary style.

Best Practices

• Simple, comma-separated sentences
• Avoid complex physical movements (bouncing balls, running)
• Avoid specific numbers ("10 sensors" → "multiple sensors")
• Always include camera language and lighting
• Use quality terms: "cinematic", "professional", "documentary style"

Kling Generation

from app.fal_helper import FalVideoGenerator

result = await FalVideoGenerator.generate_video(
    prompt="Your Kling-formatted prompt",
    model="kling_pro",  # or kling_master, kling_standard
    duration=10,
    aspect_ratio="9:16"
)

if result.get("success"):
    video_path = result["video_path"]

Environment Variables

FAL_API_KEY=your_fal_api_key    # For Kling

For more details, see prompt-examples.md and technical.md.