PROJECT STUDY · FINAL-YEAR PROJECT · 2022

From curated data to an imagined character

I trained a 512 × 512 anime StyleGAN2 model from a dataset I curated. This page summarizes what the presentation documents and separates it from the recovered model that runs in the browser.

The reported benchmark is from the final-year presentation. It is not independently reproduced here, and the browser demo does not retrain the model.

Sixteen anime images shown as generated samples in the final-year
project presentation. — Presentation sample grid, slide 22. The presentation reports FID 7.73; its evaluation logs and full protocol are not available here.

What the project records

I curated an anime illustration dataset, then trained a 512 × 512 model using Diffusion-StyleGAN2 with DiffAugment.
The presentation specifies 512-dimensional Z/W latents, batch size 32, R1 gamma 5, and zero path-length regularization and style-mixing probability, using an RTX 3060.
I explored style mixing and GANSpace edits, then used generated images as inputs to Waifu Diffusion. The StyleGAN-NADA examples used a separate pretrained model.

Style-mixing matrix from the presentation comparing edits across
groups of synthesis layers. — Slide 24 associates layers 1–6 with overall style, 7–10 with clothing type, and 11–16 with colour in this model. These observations are not universal layer guarantees.

What the evidence cannot settle

The presentation reports FID 7.73 and includes generated samples, but does not provide the evaluation logs, the full evaluation protocol, or a controlled stability comparison. I no longer have the original checkpoint hash, conversion tool and version, or exact noise settings, so I cannot establish that the recovered browser export is byte-for- byte the checkpoint behind that reported score.

The orbit at the top of the homepage is a pre-rendered walk through recovered model outputs, not a recording of training. The live demo runs inference from that recovered export; it does not reproduce training or rerun the benchmark.

See the presentation evidence · Try the recovered model · Public project record and abstract