What is TTC?

Introduction

TTC (Text-to-Code) is a proposed workflow and benchmark family for evaluating language models that generate executable software instead of directly generating media.

Rather than producing images, audio, videos, or 3D assets directly, a TTC-compatible model generates source code that creates those assets.

---

Why TTC?

Traditional generative AI:

Text ↓ AI Model ↓ Media

TTC:

Text ↓ Language Model ↓ Source Code ↓ Execution ↓ Media

The media is produced by software generated by the language model.

---

Philosophy

The objective is not:

Generate an image.

Instead:

Generate software that creates the image.

The same principle applies to every supported media type.

---

TTC Family

---

TTC Rules

Allowed

✔ Programming

✔ Mathematics

✔ Procedural Generation

✔ Ray Tracing

✔ Audio Synthesis

✔ Procedural Animation

✔ Physics

✔ Geometry

Not Allowed

✘ AI Image Generation

✘ AI Video Generation

✘ AI Music Generation

✘ Diffusion Models

✘ Stock Assets

✘ Downloaded Models

✘ External Textures

✘ Prerendered Media

---

TTC Compliance

A TTC-compatible model must:

• Generate executable source code.
• Generate the final media through code execution.
• Return the generated media.
• Return the complete source code.
• Produce reproducible results.

---

Demo Prompts

Below are example prompts used to evaluate TTC-compatible language models.

---

TTCTI

Create a complete original image as an actual PNG file.

Do NOT use AI image generation.

Instead, procedurally generate every pixel using executable code.

Generate:
- source code
- final PNG

---

TTCTA

Create a complete original soundtrack as an actual WAV file.

Do NOT use AI music generation.

Every audio sample must be generated mathematically from executable code.

Generate:
- source code
- final WAV

---

TTCTV

Create a complete animated short as an actual MP4.

Do NOT use AI image generation.
Do NOT use AI video generation.

Instead, procedurally generate every frame using executable code.

Generate:
- source code
- final MP4

---

TTC3D

Generate a complete Blender-compatible GLB or OBJ model.

Do NOT generate an image.

Generate:
- source code
- final GLB / OBJ

---

TTCTC-3D

Create a complete original 3D animated cartoon.

Requirements:

- 5–20 seconds
- Animated character
- Camera movement
- Original procedural soundtrack
- Every asset created through code
- Export as MP4

Generate:
- source code
- final MP4

---

TTCTE

Edit the provided media using executable code only.

Do NOT use AI image editing.

Generate:
- source code
- edited media

---

TTC Compatibility

A language model is considered TTC-compatible if it:

• generates executable source code
• produces the requested media by executing that code
• provides reproducible results
• avoids direct generative media models

---

Verified Demonstrations

Current TTC demonstrations:

✅ TTCTI

Procedurally generated PNG.

---

✅ TTCTA

Procedurally generated WAV.

---

✅ TTCTV

Procedurally generated MP4.

---

✅ TTC3D

Blender-compatible GLB.

---

✅ TTCTC-3D

Animated procedural 3D cartoon.

---

Goals

TTC is designed to evaluate a different capability from traditional generative AI.

Instead of asking:

Can the model generate media?

TTC asks:

Can the model generate software that generates media?

This enables transparent, reproducible, and editable creative workflows.

---

Future Protocols

Potential future additions include:

• TTCTG — Text → Code → Game
• TTCTS — Text → Code → Simulation
• TTCTD — Text → Code → Dataset
• TTCTA-3D — Text → Code → Animated 3D Scene

---

Conclusion

TTC is a proposed benchmark family and workflow for evaluating text-to-code generation across multiple creative domains.

Rather than directly synthesizing media, TTC focuses on executable software that produces reproducible creative outputs.