Home / GANs

GANs

Image models trained only on my own pictures

I treat these tools as material: train them on your own pictures, find where they fail, build the tool you need when it doesn't exist, and keep a hand on the final object.

The kernel

May – June 2026 · one GAN, 200,000 steps

Six grids from noise to the finished model

One model learning, left to right and top to bottom: pure noise at sample 0, blobs at 25, outlined creatures at 100, the archive surfacing at 600, the finished base model, then one Spell's own model.

A GAN (generative adversarial network) is two programs trained against each other: one makes images, the other tries to tell them from the real ones. Over time the maker gets good enough to fool the judge.

The archive9,839 of my own images: photographs, posters, collages and paintings, pulled from the photo pool, the collage workspace and the archive.
The trainingA lightweight GAN at 128 × 128 pixels, batches of 8, 200,000 steps. It ran for about 80 hours on my own computer.
One model per SpellEach Spell then got its own copy, trained 4,000 more steps on only that Spell's pictures (June 14 to 16, 2026), from 472 pictures for the smallest to 4,000 for the largest.
Into the workThe frames never resolve into anything representational; they stay in that unseen place, always still forming. Some become a starting layer I build a painting on top of.

Each Spell's model

64 samples per grid

Addressed to No One model samples — **Addressed to No One** · trained on 4,000 of its pictures

AI: The Instrumented Face model samples — **AI: The Instrumented Face** · trained on 642 of its pictures

AI: The Machine Re-Stages the Empty Hour model samples — **AI: The Machine Re-Stages the Empty Hour** · trained on 472 of its pictures

AI: The Obedient Silhouette model samples — **AI: The Obedient Silhouette** · trained on 1,913 of its pictures

Discarded Habitation model samples — **Discarded Habitation** · trained on 1,121 of its pictures

Permission to Pass model samples — **Permission to Pass** · trained on 4,000 of its pictures

Promise Outlives the Hour model samples — **Promise Outlives the Hour** · trained on 3,059 of its pictures

Screen Where the Face Should Be model samples — **Screen Where the Face Should Be** · trained on 3,894 of its pictures

The Detachable Face model samples — **The Detachable Face** · trained on 2,413 of its pictures

The Frame Around the Frame model samples — **The Frame Around the Frame** · trained on 1,258 of its pictures

The Wall Keeps Its Drafts model samples — **The Wall Keeps Its Drafts** · trained on 4,000 of its pictures

Waiting as the Whole of the Job model samples — **Waiting as the Whole of the Job** · trained on 4,000 of its pictures

Early training

Frames from the first hours of training, before the model had learned anything. My website uses these frames on its production-stage page.

A painting planned from training time

plan · agreed August 2026

A planned 60 × 60 inch painting: one Spell's 8 × 8 sample grid, screen-printed in six layers, each layer a different moment of training (samples 0, 25, 100, 600, 13,333 and 13,599), with a thin transparent white veil between layers. Six 24 × 24 inch paintings would be cropped from the same grids. Inks build from warm grey to two inks to CMY to CMYK.

The plan, to scale.

Mockup: the grid under a 50% white veil.

LoRAs on my own pictures

April 2026 · Stable Diffusion 1.5

A LoRA is a small add-on that teaches an existing image model one new look. I trained two on my own computer: amory_aesthetic on 478 of my own captioned images (1,200 steps, about 34 minutes) and z0glyphs on 240 synthesized glyph images. These are test outputs.