README.md

metadata

license: apache-2.0
pipeline_tag: image-generation
language: []
tags:
  - gan
  - dcgan
  - image-generation
  - logos
  - from-scratch
  - toy
library_name: torch
metrics:
  - discriminator-loss
  - generator-loss

logo-gan — a tiny DCGAN that learns to draw company logos

A small Deep Convolutional GAN trained from scratch to generate 64×64 RGB company logos. Built for a request in Compactbot/model-requests#1.

This is a toy / experiment, not a production image model. It is published mainly as a small, honest, reproducible build.

What it is

• Architecture: standard DCGAN.
  • Generator: Linear(100 → 512·8·8) + 3× up-conv (512→256→128→64) + final 3×3 conv, BatchNorm, ReLU, Tanh out.
  • Discriminator: 4× down-conv (64→128→256→512→1) + LeakyReLU(0.2).
• Parameters: 8,832,708 (generator 6,066,179 + discriminator 2,766,529). The model.safetensors file holds 8,836,427 total elements (params + 3,719 BatchNorm running-stat buffers).
• Latent: 100-dim Gaussian.
• Stability tricks: label smoothing (real=0.9, fake=0.1) + R1 gradient penalty (λ=10) on the discriminator. A vanilla run without these mode-collapsed (d→0.0000, g→13.9) and was discarded; this is the stable variant.

Training

• Data: 400 real company logos, resized to 64×64, from four public Hub datasets (100 each): samp3209/logo-dataset (bliptest), taniya/Logo_mark, taniya/Logo_symbol, taniya/Logo_type.
• Hardware: NVIDIA RTX 5090 (32 GB), ~5 min for 12,000 steps.
• Optim: Adam, generator LR 2e-4, discriminator LR 4e-5, betas (0.5, 0.999), batch 128.
• Schedule: 12,000 steps, checkpoint + sample grid every 2,000 steps.
• Loss curve: d settled to ~0.73–0.80, g rose to ~1.9–2.4 over training. No collapse.

Honest caveats (read this)

• I cannot visually verify the samples. The training environment produces PNG grids, but I have no way to look at them. The evidence that this is a working (non-collapsed) GAN is quantitative only:
  • discriminator and generator losses held in a healthy band throughout (no collapse to a single mode);
  • sample color entropy 4.91 bits / 2,979 unique colors (real data: 4.23 bits / 2,805) — comparable diversity, not a single repeated tile;
  • per-channel std ≈ 0.29–0.31 (real: 0.32) — full-color, not grayscale.
• That is not the same as "these look like logos." With 400 logos and 8.8M params the model can learn the statistics of logos (bright background, a central colored mark, some letterforms) but it will not faithfully reproduce any specific real logo. Treat samples_final.png as "what the model thinks a logo looks like," not as generated brand assets.

Files

• model.safetensors — generator + discriminator weights (final checkpoint, 55 tensors).
• samples_final.png — 64 fixed-latent sample grid from the final generator.
• train_gan_v2.py — the exact training script (self-contained, PyTorch).
• manifest.json — the 400 source logos (dataset + filename) the model trained on.

Reproduce

# needs: torch, numpy, pillow
# put logos64.npy (N,3,64,64) float32 in [0,1] at ./logos/logos64.npy
python3 train_gan_v2.py --steps 12000 --batch 128 --seed 7

Set seed 7 to reproduce the exact weights in model.safetensors.