---
title: LLM Cinema
emoji: 🎬
colorFrom: green
colorTo: gray
sdk: gradio
sdk_version: 6.16.0
app_file: server_app.py
pinned: true
short_description: A small LLM writes, draws & directs original ASCII films
thumbnail: https://conductorailabs-llm-cinema.hf.space/static/share.png
tags:
  - build-small-hackathon
  - thousand-token-wood
  - gradio
  - off-brand
  - field-notes
  - off-the-grid
  - llama-cpp
  - open-trace
---

# 🎬 LLM Cinema

A small model — **Cinema**, a fine-tuned **MiniCPM3-4B** — that **writes, draws, and
directs ORIGINAL short films in ASCII**, and plays them **right here in the page**,
streaming frame by frame. No video, no sound. Just imagination, one frame at a time.

**Two ways to run it.** This hosted Space generates live via a cloud endpoint so you can try it
instantly in the browser. It also ships as a fully **off-the-grid** build that runs the 4B model
**locally through llama.cpp** — no cloud, no API — see [`LOCAL_LLAMACPP.md`](LOCAL_LLAMACPP.md).

From a single concept it writes a 9-beat, ~90-second film: it invents the story, draws
every character in ASCII, picks the terrain (water shimmers, grass, snow, lava…), the
weather and time of day, camera moves (pan / push / shake), and a feeling for each
character — then stages and plays it on a little charcoal terminal.

## 🏆 Hackathon submission

**Track:** 🍄 An Adventure in Thousand Token Wood — something delightful that wouldn't exist
without AI. Here the model is *load-bearing*: it writes the story, draws every character, and
directs the staging. The whole thing runs on a 4B model (MiniCPM3-4B — under the ≤32B cap, and
under the ≤4B Tiny Titan bar).

- **Live Space:** https://conductorailabs-llm-cinema.hf.space
- **Demo video:** [▶ watch](https://huggingface.co/spaces/build-small-hackathon/llm-cinema/blob/main/media/llm_cinema.mp4) · [direct mp4](https://huggingface.co/spaces/build-small-hackathon/llm-cinema/resolve/main/media/llm_cinema.mp4)
- **Social post:** https://www.linkedin.com/posts/jeff-middleton-10354b5_huggingface-llm-ai-ugcPost-7472422889783291904-7c9j
- **Builder:** [ConductorAILabs](https://huggingface.co/ConductorAILabs) — registered + joined the build-small-hackathon org.

**Bonus badges**

- 🎨 **Off-Brand** — fully custom `gr.Server` frontend (cinema-hall backdrop, ASCII projected
  through the screen, reactive lighting), well past the default Gradio look.
- 🔌 **Off the Grid** — can optionally run **fully local** (no cloud APIs) via llama.cpp; see [`LOCAL_LLAMACPP.md`](LOCAL_LLAMACPP.md).
- 🦙 **Llama Champion** — the model runs through the **llama.cpp** runtime ([`serve_llamacpp.py`](serve_llamacpp.py)).
- 📓 **Field Notes** — see [`FIELD_NOTES.md`](FIELD_NOTES.md).
- 📡 **Open Trace** — the agent's scrubbed, secret-free build trace: [`traces/AGENT_TRACE.md`](traces/AGENT_TRACE.md).

## Try it

- **🎬 Make a film** — type any concept (or hit *Surprise*) and watch Cinema write, draw,
  and play an original film live.
- **🍿 Gallery** — hand-picked films that replay instantly, even before the model wakes.
- **🔁 Replay** the current film · **⛶ Fullscreen** (mobile) for a cinema view.

## How it works

- **Cinema** (LoRA-fine-tuned MiniCPM3-4B) emits a strict-JSON "movie script": title,
  logline, and shots (narration, cast, action, setting, camera, mood, dialogue).
- A from-scratch ASCII engine resolves each character (a sprite library + a draw-on-demand
  model for anything new), composes the scene with z-ordered occlusion, textured floors and
  set-dressing (sun/moon, clouds, stars, trees), then animates and subtitles it.
- A custom `gr.Server` frontend streams frames over SSE onto a cinema-hall backdrop, with
  reactive lighting that mirrors the on-screen action.
- **Optionally**, the director model can run **entirely on your machine** through the **llama.cpp**
  runtime (GGUF) — no cloud APIs. (This hosted Space uses a cloud endpoint for the live demo.)

Built for the **Build Small Hackathon** · the model & training pipeline stay private; this
app is the demo.

## Run it — off the grid (primary)

Everything runs on your own machine. The 4B model serves locally via **llama.cpp**, exposing an
OpenAI-compatible endpoint that the app points at — no cloud, no keys.

```bash
pip install -r requirements-local.txt
# serve the local GGUF model (OpenAI-compatible, llama.cpp) on :8000
MODEL=./models/cinema-q4_k_m.gguf python serve_llamacpp.py
# then, in another shell, point the app at it and open the cinema:
export CLAUDEMOVIES_LLM_URL=http://localhost:8000/v1
export CLAUDEMOVIES_LLM_KEY=local         # any non-empty string; not checked
export CLAUDEMOVIES_LLM_MODEL=cinema
python server_app.py
```

Full setup (GGUF conversion + quantization, llama.cpp build) is in
[`LOCAL_LLAMACPP.md`](LOCAL_LLAMACPP.md). The **Gallery** plays even with no model running —
films are pre-written — so the page always shows something instantly.

### Hosted demo (optional fallback)

The same app can point at any hosted OpenAI-compatible endpoint via Space **secrets**
(`CLAUDEMOVIES_LLM_URL` / `_MODEL` / `_KEY`) — nothing is committed to the repo. This is just a
convenience for the live link; the off-grid local run is the real thing.