The grey aliens from Stargate SG-1 — and, since v4, a good deal more than that: their ships, their rooms, their consoles and crystals, the armoured Vanir from Atlantis, and the ancestor found in stasis.

 

What it covers

The beings. Asgard at any distance — the dome skull, the black almond eyes, the slender torso and long thin limbs, and the head-to-body ratio the base model gets wrong. Close-ups hold. This is the main reason the LoRA exists: without it an Asgard face falls apart at close range and you have to keep them small, distant or in silhouette.

The Vanir. Two looks, both trained: the armoured suit with the lamp-lit helmet, and an Asgard carrying the head and chest pieces without the suit.

Their places. Corridors, laboratories, halls, the high council chamber, a ship bridge, an airlock, the city seen from above and from among its buildings.

Their things. Consoles, control stones, control crystals, the computer core, wall monitors, stasis cylinders, regeneration pods, the transport beam, Norse runes, and Thor's Hammer on its rock ledge.

Their ships. Three named classes — Bilskirnir, O'Neill, and the small science vessel — each callable by name.

 

Quick start

asgard is the main trigger. Put it in every prompt that has a being in frame. Strength 1.0, with one exception: asgard ancestor wants 0.7. That is not a prompt problem, it is the dial — at full strength the 286 asgard images outvote the 13 ancestor ones and pull him back to an ordinary Asgard.

The other 24 trigger words are motifs, not modifiers. They work on their own; you only need asgard when a being is in the frame. Every one of them is listed because it was tested and held.

Base model: Krea 2 — trained on Raw, generated on Turbo. Both bf16 and fp8 are attached; they produce the same images, so take fp8 if you want half the download.

 

What it does well, and what it will not do

Composition stays free. A wide shot stays a wide shot, and the LoRA does not drag everything toward a portrait. Foreign settings and styles come through cleanly — beach, kitchen, oil painting, black and white, a dancing pose all worked on the first try.

Three things it cannot do, each tested until it was clear rather than assumed:

A shield wrapped around a whole ship — seven wordings, none carried. Two different ship classes in one frame — four attempts, always a hybrid. Damage or destruction of any kind — the vocabulary simply is not there.

One habit worth knowing: light your interiors. Say bright interior light or the model pulls indoor scenes toward the dark.

 

Where the detail is

The full prompting guide sits in the v4 version notes below — every trigger word with its exact wording, the phrasings that work and the ones that quietly fail, and the reasoning behind each. It is long because it is measured rather than guessed: 413 generated images rated by eye, one at a time.

An INSTRUCTIONS.txt is attached as well. It is Markdown despite the extension — rename it to .md and hand it to an LLM, and it will write prompts for this LoRA that work.

 

Versions

v4 is the current one. A rebuild, not a touch-up: 344 captioned images instead of 68, and the coverage above instead of the beings alone.

v1 stays up for reproducibility only. If you made images with it and want them back exactly, use v1. For anything new, take v4.

 

Please note

Fan project, not affiliated with or endorsed by MGM. Unlike a character LoRA this one depicts no real person — it is a creature and technology design.

Some images in the gallery also use Stargate and DHD (SG-1) by lamarsha for the gate itself. Credit where it is due — the gate is lamarsha's work, not mine.