In my novel Solarum: The Continuance Crisis, the people of Isla Buena live in sea hulls: spinning drums ten kilometres across, moored near Saturn, where a shallow tropical sea runs all the way round the inside. Stand on the shore and the sea doesn't stop at a horizon. It climbs up in front of you, islands and reefs and all, and hangs over your head.

What it does: puts you at eye level in a stilt village, on a white beach, on a dock or in a boat, with the sea rising behind and hanging overhead.

How to use it

• Trigger: talvori (v2.0). v1.0 used jsrise; if you’re on v1.0, use that instead.

• Weight 1.0 for v2.0 (0.9 to 1.1); 1.1 for v1.0.

• Start from: talvori, a wooden pier with moored painted boats, the sea curves up into the sky and hangs overhead, islands and reefs above

- Always add: the sea curves up into the sky and hangs overhead, islands and reefs above

- Best foregrounds: stilt villages, beaches, boats and boat decks, lighthouses, harbours. Works at dusk too.

- Negative: top-down, map, whirlpool, tunnel, cartoon, 3d render, text, letters

What it can't do (yet)

v1.0’s trigger jsrise alone gives a sunrise (“rise” is a real word; my mistake). v2.0’s talvori alone gives a sea scene, but it only climbs once you add a subject and the climb phrase.

• v1.0: markets and towns drift toward thatched stilt huts (v2.0 keeps them).

• v1.0: at 0.9 it keeps your subject but often loses the climb; at 1.1 it climbs but pulls toward its own look. v2.0 holds both at 1.0.

• The sea overhead is painterly and softer than the foreground.

How it was made

A 3D sea hull built in Python (no 3D software), rendered from 30 eye-level camera spots. SDXL painted over each view, repainted only the near band at eye level, then blended the whole picture. 38 pictures (v1.0); 87 for v2.0, adding harbours, rocky shores, boat decks and markets, captioned only with what's in the foreground, so the trigger carries the climbing sea. Full story: Building Solarum, part 8 (coming soon). No other people's art, nothing from Flux / Google / OpenAI.

Training: Juggernaut XL v9 base, kohya sd-scripts, LoRA dim 32 / alpha 16, UNet only, 1024 px buckets, AdamW8bit, cosine, LR 1e-4, 15 repeats, 6 epochs (858 steps), released epoch 6. v2.0: 87 pictures, 10 repeats, released epoch 4 (872 steps).

Made with Claude (Anthropic): Claude Opus 5.5 in Cowork wrote the 3D renderer and planned the method; Claude Code with Claude Sonnet 5.5 on my desktop ran the renders and training.

More: https://signalofsolarum.substack.com · https://www.deviantart.com/signalofsolarum