FaceForge DCGAN-64 — FFHQ
Model description
FaceForge DCGAN-64 is an educational, unconditional PyTorch DCGAN that creates
64 × 64 synthetic face-like images from 128-dimensional random latent
vectors. It is the project baseline, not a production portrait model, identity
system, or image editor.
- Developed by: Tarek Amr
- Model type: Unconditional DCGAN generator
- Language: English documentation; no text input
- License / terms: mit
- Code: https://github.com/tarekamr737/Realistic-Human-Face-Generation-using-GAN
Files
| File | Purpose |
|---|
generator_best.pt | Generator state dictionary; load only through the matching FaceForge code. |
model_metadata.json | Architecture contract, preprocessing settings, and training information. |
Do not upload FFHQ image files, an R3GAN checkpoint, generated samples that
could be confused with real people, or arbitrary Python pickles to this model
repository.
Training
- Dataset: FFHQ, obtained separately under its applicable terms.
- Training resolution: 64 × 64 RGB; source files were 512 × 512 and were
resized during loading.
- Architecture: DCGAN generator with a 128-D latent input. Use the exact
latent_dim and feature_maps stored in model_metadata.json.
- Objective: BCE-with-logits adversarial objective with Adam optimizers.
- Run: [epochs, batch size, hardware, wall-clock time, date]
- Preprocessing: [confirm resize, crop policy, and normalization from metadata]
Evaluation
Record only measured results from the exported evaluation.json:
| Metric | Result | Notes |
|---|
| Generator loss | [value] | Training diagnostic, not a quality score. |
| Discriminator loss | [value] | Training diagnostic, not a quality score. |
| Diversity diagnostic | [value] | Pixel-space heuristic, not a realism metric. |
| FID | Not measured | Do not replace with an estimate. |
Inspect fixed-noise grids for artifacts and collapse. Losses, a small sample
grid, and pixel diversity do not demonstrate realism, fairness, privacy, or
fitness for a real-world deployment.
Intended use
Suitable only for education, controlled GAN demonstrations, and comparisons
with stronger pretrained models. All outputs must be visibly labelled
synthetic.
Out-of-scope use
Do not use this model for face recognition, identity inference, impersonation,
deception, profiling, biometric decisions, or claims that a generated image
depicts a real person. It should not be used as evidence of fairness, privacy,
or demographic representation.
Limitations and bias
This is a low-resolution baseline trained on a limited face dataset. Outputs
can be blurred, artifacted, repetitive, or unrepresentative. The model can
inherit dataset bias and may generate people-like images with unintended visual
similarities. It has not received a demographic, safety, privacy, or benchmark
evaluation.
Citation and acknowledgements
Credit FFHQ according to its official dataset documentation and terms. For the
optional visual reference in the FaceForge app, cite BrownVC R3GAN separately;
it is not this model and is not redistributed here.
