README.md

---
title: Tricket
emoji: 🎬
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 6.5.1
app_file: app.py
pinned: false
short_description: One prompt → a vertical product ad video
---

# 🎬 tricket — one prompt → a product ad video

Describe a product and tricket produces a finished **vertical (9:16) ad**:

1. **GLM-4.6** (via the HF inference router) writes a multi-scene ad script.
2. **FLUX.1-schnell** paints each scene.
3. **Kokoro** voices the narration (English or 中文).
4. **ffmpeg** cuts it together — Ken Burns motion, burned captions, voiceover — into an MP4.

All GPU work runs **on-demand on [Modal](https://modal.com)** and scales to zero when idle.

> Built for the build-small hackathon.

## Architecture

```
HF Space (Gradio, this repo)                Modal.com  (app: tricket-flux)
  app.py  ── GLM script (HF router)
          ── Model.generate   (FLUX.1-schnell, A100)  per scene
          ── TTS.synth        (Kokoro, CPU)           per scene
          ── assemble_video   (ffmpeg, CPU)  → 9:16 MP4
```

Two tabs: **🎬 Ad Studio** (the pipeline above) and **🖼 Single Image** (plain FLUX text-to-image).

## Setup (for maintainers)

1. Deploy the backend (Modal CLI authenticated):
   ```bash
   modal secret create huggingface HF_TOKEN=hf_xxx   # FLUX is gated
   modal deploy modal_app.py
   ```
2. Set this Space's secrets (Settings → Variables and secrets):
   - `MODAL_TOKEN_ID`, `MODAL_TOKEN_SECRET` — call Modal from the Space
   - `HF_TOKEN` — used by the GLM script writer (router.huggingface.co)