Media
Generate images, videos, and audio via fal.ai models over MCP with progressive quality — iterate cheap, finalize expensive. Modes: image (text-to-image + editing), video (text/image-to-video), audio (speech, music, SFX, video-to-audio).
Workflow
Applies §10.2 YAGNI (don't pay for production models until composition is locked) and §10.1 KISS (simplest model that meets the need).
- Gate: MCP — verify the fal.ai MCP server is configured. If missing, give setup instructions (key at fal.ai):
"fal-ai": {
"command": "npx",
"args": ["-y", "fal-ai-mcp-server"],
"env": { "FAL_KEY": "YOUR_FAL_KEY_HERE" }
}
- Estimate cost before generating; report the estimate for expensive runs (video especially):
estimate_cost(model_name: "fal-ai/...", input: {...}).
- Gate: ElevenLabs — for TTS via ElevenLabs, verify
ELEVENLABS_API_KEY; else fall back to csm-1b or inform the user.
- Generate progressively — cheap model for prompt iteration, lock with
seed, then switch to the production model for finals.
- Deliver — file path/URL, model + parameters, cost incurred, iteration suggestions if quality is low.
Progression: nano-banana-2 -> nano-banana-pro · seedance-1-0-pro -> veo-3 · csm-1b -> ElevenLabs.
Model Reference
| Model |
Type |
Best For |
Cost Tier |
fal-ai/nano-banana-2 |
Image |
Quick iterations, drafts, image editing |
Low |
fal-ai/nano-banana-pro |
Image |
Production images, realism, typography |
Medium |
fal-ai/seedance-1-0-pro |
Video |
Text-to-video, image-to-video, high motion |
High |
fal-ai/kling-video/v3/pro |
Video |
Text/image-to-video with native audio |
High |
fal-ai/veo-3 |
Video |
Video with generated sound, high visual quality |
High |
fal-ai/csm-1b |
Audio |
Conversational text-to-speech |
Low |
fal-ai/thinksound |
Audio |
Video-to-audio (matching sounds from video) |
Medium |
- Image params:
prompt (required); image_size (square, portrait_4_3, landscape_16_9, portrait_16_9, landscape_4_3); num_images (1-4); seed; guidance_scale (1-20, higher = more literal).
- Video params:
prompt (required); duration ("5s", "10s"); aspect_ratio ("16:9", "9:16", "1:1"); seed; image_url (source for image-to-video).
- MCP tools:
search, find, generate, result, status, cancel, estimate_cost, models, upload.
Image Editing
Nano Banana 2 with an input image for inpainting, outpainting, or style transfer:
upload(file_path: "/path/to/image.png")
generate(model_name: "fal-ai/nano-banana-2", input: {
"prompt": "same scene but in watercolor style",
"image_url": "<uploaded_url>",
"image_size": "landscape_16_9"
})
For non-MCP integrations (ElevenLabs, VideoDB), follow handlers/external-apis.md.
Common Mistakes
Using pure text-to-video when image-to-video would be more controllable; assuming fal.ai access covers ElevenLabs credentials (separate key).
Integration
- Called by: user directly,
/ai-build, ai-video-editing (Layer 5 generated assets)
- Calls: fal.ai MCP, ElevenLabs API, VideoDB API
- See also:
/ai-visual (composed visuals), /ai-slides (deck visuals), /ai-animation
Examples
/ai-media image hero for parallel agent planning blog
Iterates with nano-banana-2 (cheap), locks composition with seed, switches to nano-banana-pro for the production final, returns URL + cost.
$ARGUMENTS
1---2name: ai-media3description: Generates images, videos, and audio via AI models (fal-ai MCP): cheap iteration models, expensive production finals, cost-estimate before generation. Trigger for 'generate an image', 'create a thumbnail', 'make a voiceover', 'AI video', 'text to speech for'. Not for design composition; use /ai-visual instead. Not for animation specs; use /ai-animation instead.4---56# Media78Generate images, videos, and audio via fal.ai models over MCP with progressive quality — iterate cheap, finalize expensive. Modes: `image` (text-to-image + editing), `video` (text/image-to-video), `audio` (speech, music, SFX, video-to-audio).910## Workflow1112Applies §10.2 YAGNI (don't pay for production models until composition is locked) and §10.1 KISS (simplest model that meets the need).13141. **Gate: MCP** — verify the fal.ai MCP server is configured. If missing, give setup instructions (key at [fal.ai](https://fal.ai)):1516```json17"fal-ai": {18 "command": "npx",19 "args": ["-y", "fal-ai-mcp-server"],20 "env": { "FAL_KEY": "YOUR_FAL_KEY_HERE" }21}22```23242. **Estimate cost** before generating; report the estimate for expensive runs (video especially): `estimate_cost(model_name: "fal-ai/...", input: {...})`.253. **Gate: ElevenLabs** — for TTS via ElevenLabs, verify `ELEVENLABS_API_KEY`; else fall back to `csm-1b` or inform the user.264. **Generate progressively** — cheap model for prompt iteration, lock with `seed`, then switch to the production model for finals.275. **Deliver** — file path/URL, model + parameters, cost incurred, iteration suggestions if quality is low.2829Progression: nano-banana-2 -> nano-banana-pro · seedance-1-0-pro -> veo-3 · csm-1b -> ElevenLabs.3031## Model Reference3233| Model | Type | Best For | Cost Tier |34| --------------------------- | ----- | ----------------------------------------------- | --------- |35| `fal-ai/nano-banana-2` | Image | Quick iterations, drafts, image editing | Low |36| `fal-ai/nano-banana-pro` | Image | Production images, realism, typography | Medium |37| `fal-ai/seedance-1-0-pro` | Video | Text-to-video, image-to-video, high motion | High |38| `fal-ai/kling-video/v3/pro` | Video | Text/image-to-video with native audio | High |39| `fal-ai/veo-3` | Video | Video with generated sound, high visual quality | High |40| `fal-ai/csm-1b` | Audio | Conversational text-to-speech | Low |41| `fal-ai/thinksound` | Audio | Video-to-audio (matching sounds from video) | Medium |4243- **Image params**: `prompt` (required); `image_size` (`square`, `portrait_4_3`, `landscape_16_9`, `portrait_16_9`, `landscape_4_3`); `num_images` (1-4); `seed`; `guidance_scale` (1-20, higher = more literal).44- **Video params**: `prompt` (required); `duration` (`"5s"`, `"10s"`); `aspect_ratio` (`"16:9"`, `"9:16"`, `"1:1"`); `seed`; `image_url` (source for image-to-video).45- **MCP tools**: `search`, `find`, `generate`, `result`, `status`, `cancel`, `estimate_cost`, `models`, `upload`.4647## Image Editing4849Nano Banana 2 with an input image for inpainting, outpainting, or style transfer:5051```52upload(file_path: "/path/to/image.png")53generate(model_name: "fal-ai/nano-banana-2", input: {54 "prompt": "same scene but in watercolor style",55 "image_url": "<uploaded_url>",56 "image_size": "landscape_16_9"57})58```5960For non-MCP integrations (ElevenLabs, VideoDB), follow `handlers/external-apis.md`.6162## Common Mistakes6364Using pure text-to-video when image-to-video would be more controllable; assuming fal.ai access covers ElevenLabs credentials (separate key).6566## Integration6768- **Called by**: user directly, `/ai-build`, `ai-video-editing` (Layer 5 generated assets)69- **Calls**: fal.ai MCP, ElevenLabs API, VideoDB API70- **See also**: `/ai-visual` (composed visuals), `/ai-slides` (deck visuals), `/ai-animation`7172## Examples7374```75/ai-media image hero for parallel agent planning blog76```7778Iterates with `nano-banana-2` (cheap), locks composition with `seed`, switches to `nano-banana-pro` for the production final, returns URL + cost.7980$ARGUMENTS
arcasilesgroup