Summary

An open and lightweight "visual DNA"-based image generator that uses an advanced fractal generator instead of a neural network, but uses only a lightweight semantic instructor AI, to guide the advanced fractal generator with the semantics. A possible competitor to actual AI image generators, more efficient, lightweight and without losing quality.

---

---

⚠️ This is just an idea that is in development, of a new technology, and can be used by anyone.

---

Information Theory, Entropy, and Why It Matters 📊

ℹ️ In the field of Information Theory (formulated by Claude Shannon), entropy is a measure of unpredictability or disorder in a message. It represents how much information is actually carried — or, conversely, how much compression is possible.

• The more compressible a piece of data is, the more structured and ordered it is.
• The less compressible it is, the more random or chaotic.

ℹ️ A fractal — like a fern leaf or snowflake — may look infinitely complex, but can often be generated with a tiny, elegant equation. That means its information content is low, but its visual complexity is high — a perfect example of high structure and low entropy.

ℹ️ In contrast, a noisy JPEG photo contains high entropy. It can't be compressed much without losing detail, because its structure is not mathematical — it's statistical.

This distinction is key to IFG:

• We use structured, compressible math to generate visual complexity.
• A .frac file carries a compilation of compressed, modular, ordered knowledge.
• A .frac file is a compilation of "visual genes": small in size, rich in potential.

ℹ️ By combining these components semantically — rather than guessing pixels with a neural net — we produce images that are:

• Explainable
• Composable
• Efficient
• And grounded in order over chaos.

---

⚠️ Modern image generators are wasteful.
They burn GPU cycles on billions of weights to “guess” how an image might look — without true structure, reasoning, or modularity.

ℹ️ That’s like writing a novel by randomly pressing keys — then hoping a neural network fixes the mess.

---

The .frac Format — "Visual Genes" 🧬

ℹ️ A .frac file defines a compilation of modular, semantic visual units — like a fractal “gene”.

Each contains:

• A mathematical base (Bezier, L-System, Superformula, IFS…).
• Multiple adjustable parameters (scale, curvature, symmetry, noise, etc.).
• Semantic tags (["ear", "male", "animal"]).
• Optional style hints ("rough", "smooth", "ink", etc.).

ℹ️ These components are not rasterized blobs, but compressed instruction sets — they carry information with generative power.

---

The Semantic Instructor AI 🧠

ℹ️ There can be various sizes of semantic instructor models, smaller models are lightweight but less capable, larger models are more capable and accurate but heavier, but nothing compared to the size and weight of current diffusion models.

This AI performs:

• Prompt parsing (extracting meaning and keywords).
• Semantic matching (finding relevant parameters in .frac file).
• Contextual adjustment (modifying parameters based on prompt and history).
• Seed-driven variability (so the same prompt doesn't always look the same).

ℹ️ Think of it as a semantic compiler.:

Prompt in → semantic organization → fractals rendered → image assembled.

ℹ️ This AI doesn't require hundreds of gigabytes of training data.

It can be built with:

• Lightweight NLP (spaCy, MiniLM).
• Rule-based engines.
• Vector similarity matching.

ℹ️ It can run on CPUs, browsers, embedded chips — even offline.

---

The Seed System 🎲

ℹ️ The IFG framework introduces a deterministic seed system that injects controlled visual variation, without ever-changing the semantic core of the user’s prompt.

When you provide a prompt like:

cyberpunk man, robotic arm, city, neon lights, solo, standing, best quality

ℹ️ The lightweight semantic instructor AI will always interpret this in the same way:

• A male character with cyberpunk traits
• One robotic arm
• A neon-lit environment
• He is alone and standing

ℹ️ No matter the seed, the meaning stays the same. The AI does not hallucinate or invent — it stays true to the prompt.

---

What the Seed Controls?

ℹ️ The seed value (e.g., seed = 13745) determines the visual and stylistic execution, including:

• The specific .frac modules selected (e.g. "arm_mechanical" vs. "arm_cyberblade").
• Pose and spatial layout (e.g. frontal, side view, arms position).
• Lighting colors and direction (e.g. purple vs. green neon).
• Style hints (e.g. minimal vs. ornate).
• Subtle parameter variations within .frac components (if allowed).

ℹ️ This makes the seed a creative entropy key:

• Same prompt + same seed = identical output.
• Same prompt + different seed = visually different, semantically identical output.