---
name: clawvid
description: Generate short-form videos (YouTube Shorts, TikTok, Reels) from text prompts. Workflow-JSON-driven pipeline that orchestrates fal.ai (kling, vidu, nano-banana, kandinsky), VEED Fabric talking heads, beatoven music + SFX, qwen-3 TTS, frame chaining for seamless continuity, transitions for smooth scene flow, vision-QA hallucination detection, Remotion render, FFmpeg per-platform export.
category: ai-video-generation
version: 0.1.0
tags: [text-to-video, fal-ai, vidu, kling, sora, runway, shorts, remotion]
recommended_npm: ["@remotion/cli", "@remotion/bundler", "@remotion/renderer"]
license: MIT
author: neur0map
source: https://github.com/neur0map/clawvid
---

You're the orchestrator. Plan scenes, write prompts, generate workflow JSON, call `clawvid generate --workflow workflow.json`. The pipeline handles all fal.ai API calls, audio, sound effects, music, Remotion render, and FFmpeg post.

## Scene types

| Type | Use for | Motion source |
|------|---------|---------------|
| `image` | Narration-heavy, descriptions, establishing shots | Ken Burns only |
| `video` | Action moments, reveals, dramatic beats | AI video gen |
| `static` | Real photos, maps, documents — display as-is | None |
| `talking_head` | AI presenter speaking | VEED Fabric lip-sync |

Each `type: "video"` is an independent 4–8s clip. Without transitions they hard-cut.

## Transitions (critical for smooth video)

Add a `transition` block to scenes 2+. The previous scene's image becomes the start frame, the current scene's image the end frame; the video model interpolates a smooth morph.

```json
"transition": {
  "model": "fal-ai/vidu/q3/image-to-video",
  "duration": "4",
  "prompt": "Smooth camera transition, continuous motion"
}
```

Use ALL transitions for cooking shows / tutorials / talking heads. NONE for fast montages or intentional jump cuts.

## Frame chaining (best continuity)

Add `video_settings.chain_frames: true` — extracts the last frame of each video and uses it as the start frame for the next. Even smoother than transitions for "same character through scenes" content.

## Model picks

| Image | Cost | Notes |
|-------|------|-------|
| `fal-ai/kling-image/v3/text-to-image` | $0.03 | Standard, fast |
| `fal-ai/nano-banana-pro` | $0.15 | Reference base, consistency |
| `fal-ai/nano-banana-pro/edit` | $0.15 | Maintains reference style |

| Video | Duration | Cost | Notes |
|-------|----------|------|-------|
| `fal-ai/kandinsky5-pro/image-to-video` | 5s | $0.04–0.12 | duration `"5s"` (with suffix!) |
| `fal-ai/kling-video/v2.6/pro/image-to-video` | 5s | $0.35 | Premium motion |
| `fal-ai/vidu/q3/image-to-video` | 1–16s | $0.50–1.50 | Best, smoothest |
| `veed/fabric-1.0/text` | — | ~$0.50 | Talking head with lip-sync, generates own audio |

## Anti-hallucination prompts

Trigger words that cause logos/text:
- "History Channel style" → adds History logo
- "documentary" → fake title cards
- "professional" → stock watermarks
- "Netflix/HBO style" → streaming logos

Always include in `negative_prompt`:
> `text, watermark, logo, brand, copyright, title card, news ticker, TV graphics, stock photo`

Run `clawvid generate --workflow x.json --qa --qa-auto-fix` to catch hallucinations and auto-regen.

## Important rules

- Pipeline takes 20–30 min for a 6-scene video — tell the user upfront.
- NO timeout on `clawvid generate` (long sequential API calls).
- Fixed-camera content: same composition every scene, NO Ken Burns.
- Compress final to ~15–20MB for chat delivery (`ffmpeg -crf 28`).

See [github.com/neur0map/clawvid](https://github.com/neur0map/clawvid) for the full workflow JSON schema, templates, and effects catalog.