<pre><code>This LoRA is trained on **195 photographs of the real objects** — museum pieces, not other people's art — to teach the
six named Viking Age art styles and the things they were carved and cast into.

## Triggers

**`Viking`** is the trigger word. Then name the style the way you'd say it out loud:

| write this | you get |
|---|---|
| `Oseberg style` | gripping beasts, dense interlace, deep relief in dark oak — the ship's stem, the animal-head posts |
| `Borre style` | the ring-chain, gripping beasts with cat-like masks, gilt bronze |
| `Jelling style` | S-shaped ribbon beasts, double contour, pigtails and lip lappets |
| `Mammen style` | pelleted bodies, foliate tendrils sprouting from the beast |
| `Ringerike style` | lobed pear-shaped leaf tendrils and tight spiral curls, the great beast; the weather vanes |
| `Urnes style` | ONE slender ribbon beast looping through itself in figure-eight loops of thin and thick bands |

Add `Viking Age wood carving` / `metalwork` / `carved stone` to steer the material. Objects have their own phrases:
`the Gjermundbu helmet`, `the Mästermyr chest`, `Viking Age key`, `Viking Age runestone`, `Viking Age picture stone
from Gotland`, and `Viking Age sword blade, close on the inlaid inscription` — which will give you a genuine flush
`+VLFBERH+T` inlay rather than modern engraving.

## Settings

Start at **1.0**. Styles hold from about 0.6; below that the base model's Celtic habits come back. Above 1.5 it starts
favouring surface texture over composition — good for a carved band filling the frame, bad for a whole object.

Trained on Krea 2 **Raw**, run on **Turbo** — the usual paradigm, and every example here was generated on Turbo at 8
steps so what you see is what you get.

**It has never seen a scene.** 195 museum object photographs: no people, no landscape, no weather. Scenes still work
well — ask the base model for the scene and let this steer the ornament on it (a shield wall whose shields carry real
Borre ring-chain, a longship whose stem post carries a real Urnes band) — but drop to about 0.9 and give the prompt its
own camera and lighting.

**Everything it learned is excavated.** Rust, verdigris, chipped edges, missing lobes. The captions name the condition
on nearly every image, so you can steer it — "bright, newly forged, crisp and unworn" has something to push against —
but the default is eleven hundred years in the ground.

## Dataset

195 photographs from Wikimedia Commons, hand-picked from roughly 900 candidates and **hand-captioned — no
auto-tagging**. Free licences only: CC0, public domain, CC BY, CC BY-SA. Nothing NonCommercial or NoDerivatives.
Modern reproductions excluded on sight, which took some doing — a search for "Viking sword hilt" is mostly a replica
shop.

Oseberg 35 · picture stones 27 · Urnes 19 · Ringerike 18 · runestones 13 · Borre 12 · Jelling 8 · Mammen 6 · objects 57
(Gjermundbu helmet, Mästermyr chest and tools, Ulfberht blades, keys, brooches, a merchant's folding balance).

Rank 32 / alpha 32, 1024 with aspect buckets, 2000 steps, lr 1e-4 constant, AdamW8bit, 20 epochs saved.

## Licence and credits

Shared open. 118 of the 195 source photographs are CC BY-SA, so share-alike is treated as flowing through: use it for
anything, including commercially, and pass on what you build in the same spirit.

Per-image credits — photographer, licence and source page for all 195 — are in the attached `CREDITS.md`.

Thanks to the museums that put their collections online under free licences — the Swedish History Museum,
Kulturhistorisk museum in Oslo, Riksantikvarieämbetet and the Nationalmuseet — and to the Wikimedia photographers named
in the credits. Without CC0 and CC BY museum data this model could not exist.</code></pre><p></p>

Viking Age Art — Norse Carving Styles (Urnes, Ringerike, Borre, Oseberg, Jelling, Mammen)

<p>First release. Trained on 195 hand-captioned photographs of real Viking Age objects from Wikimedia Commons — free licences only, no auto-tagging, no modern reproductions. Epoch 20 of 20; rank 32 / alpha 32, 2000 steps, lr 1e-4 constant, 1024 with aspect buckets. Trained on Krea 2 Raw, generated on Turbo.</p>

Files with SHA256: 79a7d43335d6451d4d478cce1cd40300077a745a45f8e667be2df2a469d17a31