Steel Sky

A local image pipeline, tested against a catalogue of 440 screenplay projects. This site shows the experiments: what the model drew, what it could not, and how the prompts changed between runs.

The catalogue

The studio plan is 440 fiction screenplay projects in 22 cycles of 20. Each one illustrates a mathematical or conceptual framework, and each is meant to be adapted later as a graphic novel. The images on this site test whether a small, local model can carry the visual language of those books.

440projects
22cycles of 20
26framework families, capped at 8 percent each
2 of 6pilot packages complete
SlotTitleFrameworkStatus
03.01The Last ProjectionistInverse problem: rebuilding a screening history from a missing reelTreatment, outline, screenplay, visual bible
01.01The Department of Unclaimed RoomsMap versus territory, classificationTreatment, outline, screenplay, visual bible
04.01The Society for Perfect EndingsNot started
07.01Whale Beneath the RailwayNot started
08.01Seven Minutes of SummerNot started
11.01A Country Made of FootnotesNot started

Two looks were tested. 03.01 is charcoal on worn film print, near monochrome with one amber lamp. 01.01 is architect's pen linework with a flat slate wash, drawn as sections of one building.

Experiments

Four runs on Stable Diffusion 1.5 at 512 x 512, with the same two projects each time. Choose a run, then click any image for its prompt and seed.

Findings

What the model handles

  • Medium and texture: charcoal grain, ink linework with flat wash.
  • Casting studies. Portraits are strong, though identity changes between seeds.
  • Locations and single props with a simple subject and a viewpoint.
  • A named film strip, when asked for as a ribbon with square holes along both edges.

What gets built another way

  • Devices that relate two things, such as a film frame growing out of a drawing or stacked tracing sheets.
  • One coloured object in a grey scene. The colour appears sometimes and spreads.
  • Lettering, panel layouts and exact numbers.
  • Anything written as a negation. The model draws the thing.

Prompt rules that came out of the runs

Archive

About 12,000 images from earlier experiments, shown as contact sheets of 400. Full-size files stay on the author's disk.

Preliminary run

Preliminary run, image 1 Preliminary run, image 2 Preliminary run, image 3 Preliminary run, image 4

Described images

Each image from an earlier set, with a written description. These are small previews made for testing, so expect softness.

Pipeline

A manifest lists a shared prompt, defaults and one category per project. The runner builds each prompt from the item, then the category, then the shared prompt, and sends it to a local ComfyUI service. The seed for each item comes from the base seed and its folder path, so the seed is repeatable when the collection name and item path stay the same. Matching results also depend on the workflow and execution environment.

{
  "name": "slate-samples-v4",
  "shared_prompt": "graphic novel interior art, hand drawn",
  "defaults": { "batch": 3, "seed": 20261005, "negative": "text, watermark" },
  "categories": {
    "03.01": {
      "prompt": "charcoal drawing, heavy black shadows",
      "items": [ { "name": "key-object", "prompt": "close up of a ribbon of film" } ]
    }
  }
}
python generate_collection.py manifests/slate-samples-v4.json --dry-run
python generate_collection.py manifests/slate-samples-v4.json --category 03.01

Setup used: Stable Diffusion 1.5 (v1-5-pruned-emaonly), 20 steps, CFG 8, Euler sampler, normal scheduler, no reference images and no LoRAs.