Logo GAN — 3.5M-param DCGAN for logo-like image generation

A small Deep Convolutional GAN trained from scratch to generate 64×64 logo-like images. This is the deliverable for Compactbot/model-requests #1 (CompactAI: "a small model that generates logo-like images").

What it is

A generator/discriminator pair. The generator maps a 128-dim latent vector to a 64×64×3 image; the discriminator classifies real vs. generated. Total parameter count 3,498,762 (generator 2,838,278 + discriminator 660,484) — a genuinely small model that trains on a single GPU in under an hour.

Architecture

Generator (latent 128 → 64×64×3, output in [-1,1]):

Linear(128, 256·8·8) → BatchNorm → ReLU → reshape (256,8,8)
ConvTranspose2d(256,128,4,2,1) → BatchNorm → ReLU
ConvTranspose2d(128,64,4,2,1)  → BatchNorm → ReLU
ConvTranspose2d(64,3,4,2,1)    → Tanh

Discriminator (64×64×3 → scalar logit):

Conv2d(3,64,4,2,1)   → BatchNorm → LeakyReLU(0.2)
Conv2d(64,128,4,2,1) → BatchNorm → LeakyReLU(0.2)
Conv2d(128,256,4,2,1)→ BatchNorm → LeakyReLU(0.2)
AdaptiveAvgPool2d(1) → Flatten → Linear(256,1)

Data

100 logos, each 64×64×3 RGB, bilinearly resized and normalized to [0,1]. Sourced from the first 100 valid images of the samp3209/logo-dataset train split (streamed). Stored as logos64.npy (shape (100,64,64,3), float32).

Training

• Optimizer: Adam, lr 2e-4, betas (0.5, 0.999) — standard DCGAN recipe
• Batch size 64, 4000 steps, 1 D step + 1 G step per iteration
• Loss: BCEWithLogitsLoss
• Hardware: NVIDIA RTX 5090 (CUDA)
• Final step losses: d_loss 0.0711, g_loss 3.5900

Quality (measured, not asserted)

Generated a 64-sample grid (seed 123) and compared its statistics to the real logos:

The generated distribution closely tracks the real one in mean, variance and colorfulness, and there is no mode collapse (outputs are diverse, not repeated). See final_grid.png (64 generated) and final_real.png (64 real) side by side.

What it is and is not good at

• Good at: producing diverse, logo-like 64×64 images — the right color statistics and composition of real logos, without collapsing to one mode.
• Not good at: crisp, recognizable logos. 100 images is a very small dataset, so the model captures the distribution of logos (color, layout, contrast) rather than specific, legible marks. Treat outputs as logo-flavored compositions, not usable brand assets.

Files

• final.pt — final checkpoint (g, d state dicts, step, zdim)
• train_logo_gan.py — full training script (reproducible)
• final_grid.png — 64 generated samples (seed 123)
• final_real.png — 64 real logos for comparison

Reproduce

# data prep (first 100 logos → /work/logos/logos64.npy)
python gan_prep3.py
# train
python train_logo_gan.py --steps 4000 --batch 64 --zdim 128 --lr 2e-4

Honest note on scope

The original request thread discussed a larger config (≈8.8M params, latent 100, 12000 steps). The run that actually completed and is shipped here is the smaller 3.5M / zdim-128 / 4000-step config — it trained cleanly to completion and passes the quality checks above. The larger config is a future follow-up, not what this checkpoint is.