# Skill: Generate images, video, and audio via fal.ai with cost-efficient iteration

> When a project needs visual or audio assets, use fal.ai through an AI agent: prototype on the cheap fast model, lock in quality on the expensive model, and always estimate video costs before running. Applies to any project requiring original covers, clips, icons, or voiceovers.

Source: sodigi·learn — vibecoding/generim-kartinki-video · https://sodigi.io/learn/vibecoding/generim-kartinki-video

## When to use
- A site, bot, or presentation needs a cover image, banner, icon set, or background that matches a specific scene.
- You need a short video clip (5-10s) for social or a product demo without a camera crew.
- You need a text-to-speech voiceover or ambient audio for a video.
- You have a finished still image and want to animate it into a short clip.

## Core rules
- Draft on the cheap fast model (e.g., Nano Banana 2 equivalent), finalize on the quality model (e.g., Nano Banana Pro equivalent). Never start with the expensive model for ideas.
- For every image prompt, specify: scene description, lighting, style/mood, colors, and frame shape (square / landscape_16_9 / portrait_16_9). Vague prompts produce average results.
- Request 3-4 variations in one call during the draft phase; pick the best before moving to the quality model.
- Always call `estimate_cost` before running video or audio generation -- video is many times more expensive than images.
- For predictable video, generate a polished still image first, then use image-to-video. Text-to-video is less predictable.
- Use a seed to lock in randomness when fine-tuning a nearly-correct result -- same seed + adjusted prompt keeps the frame direction stable.
- AI models garble exact text embedded in images -- add any important copy as a separate overlay layer, not inside the image prompt.
- Break long video needs into 5-second scenes; long single clips are expensive and finicky.

## Procedure
1. Define the asset: type (image/video/audio), frame shape, scene, mood, colors.
2. Run 3-4 draft variations on the cheap model. Evaluate and pick the best direction.
3. For images: move the winner to the quality model for the final pass.
4. For video: first build and approve the still image, then run image-to-video with a motion description. Estimate cost before running.
5. For audio: use text-to-speech for voiceovers (CSM-1B equivalent) or scene-matched ambient sound (ThinkSound equivalent). Estimate cost before running.
6. Note the seed of any result worth keeping before making further adjustments.

## Ready-to-use prompt
```
Generate an image for [describe the use case]. Use a fast cheap model for drafts and generate 4 variations at once.

Scene: [describe what is in the frame -- subjects, setting, position]
Light: [e.g., soft morning light from the left, dramatic backlight]
Mood and colors: [e.g., warm and calm, beige and brown]
Style: [e.g., photorealistic, flat illustration, editorial]
Frame shape: [square / landscape_16_9 / portrait_16_9]

Estimate the cost before generating.
```

For image-to-video:
```
Take my finished image and make a [N]-second clip. Motion: [describe movement -- camera pull-back, steam rising, light flicker]. Frame shape: [16:9 / 9:16]. Estimate cost first, then run.
```

## Pitfalls
- Vague prompt ("make a coffee picture") -- produces generic stock-like output; always specify scene, light, style, colors.
- Starting with the expensive quality model for drafts -- one botched draft costs as much as ten cheap ones.
- Running video without estimating cost first -- video billing can be a surprise; always check price before hitting generate.
- Generating video from pure text when you need a specific visual -- perfect the still image first, then animate it.
- Expecting accurate text to appear inside generated images -- AI scrambles letters; add text as a post-process overlay.
- Generating a 60-second clip in one request -- expensive and hard to iterate; cut into 5-second scenes instead.