Raul OS / CarDreamer

Bachelor's thesis · University of Jaén · 27 Jun 2023

CarDreamer: cars dreamed by a neural network.

I designed, trained, and compared generative models to produce car images. This version preserves the evidence, explains the limitations, and never pretends the original run can still be reproduced.

Historical FID · feature 192

Loading verified series…

Historical GAN FID progressionA 1,001-observation curve with a 3.5374 minimum and sharp degradation after index 844.

Visual evidence

Recognisable, with limited diversity.

The selected thesis epoch produced recognisable vehicles. Several samples also repeat structure, colour, and background: visible evidence of insufficient diversity.

Grid of twenty-five cars generated by the CarDreamer GAN, with several repeated vehicles and backgrounds.
Thesis figure 5.10, attributed epoch 843. Original work. Extracted without signatures or personal metadata.
SIGNAL 01

Isolated FID minimum

3.5374 is a historical point under one protocol, not a universal score.

SIGNAL 02

Later collapse

After the minimum, FID and generator loss deteriorate while discriminator loss reaches zero.

SIGNAL 03

Validation leakage

The recovered code mixes train and test; evaluation is not independent.

What I built — and what I would correct today.

The work covered research, Stanford Cars preparation, architecture design, training, metric instrumentation, and visual analysis. The recovered material confirms GAN and VAE implementations plus a TransGAN exploration.

Methodological reading

  • Historical FID uses the 192-feature layer and small subsets.
  • The GAN loss series contain an initial `NaN`.
  • No original weights or per-epoch outputs remain; this page therefore uses replay.

Separate modern reproduction

A current version would lock seeds and environment, preserve an independent test split, calculate FID 2,048 and KID across multiple runs, and publish density/coverage alongside failed samples.

Sources

GAN · VAE · TransGAN · FID