PixelModel 🖼️

A neural network where the weights are the image.

📌 What is this?

model.png is not a picture — it is the model.

Every pixel encodes neural network weights. At inference, the PNG is decoded into weight matrices forming a tiny MLP. The prompt is embedded into a vector, and the model generates a 32×32 image.

Training directly optimizes pixel values via gradient descent until the PNG becomes the model itself.

---

🎨 Weight Encoding

• R channel → weight magnitude (0–255 → 0.0–1.0)
• B channel → weight sign (<128 = negative, ≥128 = positive)
• G channel → unused / reserved

---

🧠 Architecture

prompt string
  → char embedding → 32-dim vector
  → W1 (64×32)  → tanh
  → W2 (64×64)  → tanh
  → W3 (3072×64) → sigmoid
  → reshape → 32×32×3 image

All weights live inside model.png.

---

📦 Standard weights (safetensors)

model.png is the canonical model — training writes to it directly, and it's what makes PixelModel PixelModel. For tooling that expects standard weight files, the same 3 matrices are also exported as model.safetensors (202,752 parameters total, no bias terms):

python convert_to_safetensors.py            # model.png -> model.safetensors
python convert_to_safetensors.py --model model.png --out model.safetensors

Re-run this after training if you retrain into a new model.png — model.safetensors doesn't update itself.

Parameter count is verifiable two ways without running any code: config.json (total_parameters: 202752, full per-layer breakdown) and the safetensors file's own header metadata (total_parameters, param_breakdown, has_bias, text_encoder_parameters, vae_parameters — all 0 except the MLP itself).

---

🧪 Dataset vs Outputs

Target	Output

---

📁 Files

model.png                   ← THE MODEL (64×3200 px)
model.safetensors           ← same weights, standard format (generated, see below)
config.json                 ← architecture + parameter-count metadata
main.py                     ← inference, loads model.png
INFERENCE.py                ← inference, loads model.safetensors
convert_to_safetensors.py   ← model.png -> model.safetensors
train.py                    ← training
model.py                    ← architecture
dataset/
  red.png
  red.txt       ← prompt: "red"
  ...

---

⚙️ Usage

python train.py
python train.py --epochs 500 --lr 0.05

python main.py "red"
python main.py "a cat" --out cat.png --scale 8

# equivalent, but loads model.safetensors instead of model.png
python convert_to_safetensors.py
python INFERENCE.py "a cat" --out cat.png --scale 8

main.py and INFERENCE.py produce byte-identical output for the same prompt — they're the same architecture and weights, just loaded from different files.

---

📊 Tips

• 6–20 samples are enough
• Simple patterns converge fastest
• 200–500 epochs typical
• Loss < 0.001 is strong for toy datasets

---

It’s a toy. It’s not useful. But it works.

Bench Labs · Simple, Reliable, Open sourced