Views
No views yet
|
🏆 BenchLabs Leaderboard •
🌍 Tiny-T2I Leaderboard
|
model.png is not a picture — it is the model.1prompt string
2 → char embedding → 32-dim vector
3 → W1 (64×32) → tanh
4 → W2 (64×64) → tanh
5 → W3 (3072×64) → sigmoid
6 → reshape → 32×32×3 imagemodel.png.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):1python convert_to_safetensors.py # model.png -> model.safetensors
2python convert_to_safetensors.py --model model.png --out model.safetensorsmodel.png — model.safetensors doesn't update itself.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).| Target | Output |
|---|---|
![]() | ![]() |
![]() | ![]() |
![]() | ![]() |
![]() | ![]() |
![]() | ![]() |
![]() | ![]() |
1model.png ← THE MODEL (64×3200 px)
2model.safetensors ← same weights, standard format (generated, see below)
3config.json ← architecture + parameter-count metadata
4main.py ← inference, loads model.png
5INFERENCE.py ← inference, loads model.safetensors
6convert_to_safetensors.py ← model.png -> model.safetensors
7train.py ← training
8model.py ← architecture
9dataset/
10 red.png
11 red.txt ← prompt: "red"
12 ...1python train.py
2python train.py --epochs 500 --lr 0.05
3
4python main.py "red"
5python main.py "a cat" --out cat.png --scale 8
6
7# equivalent, but loads model.safetensors instead of model.png
8python convert_to_safetensors.py
9python INFERENCE.py "a cat" --out cat.png --scale 8main.py and INFERENCE.py produce byte-identical output for the same prompt — they're the same architecture and weights, just loaded from different files.