Concepts and the first mesh

This stage used native proportions to revise prompts and inspect the first generated mesh. FLUX was used for those early concepts; later work uses imagegen concepts while retaining Meshy for base geometry.

The later model and motion are shown in the complete animation.

As of September 9, 2026: creature mod 0.12.1 0.12.2 covers all fourteen Necropolis creatures. Measurements here retain their delivery-version scope; see mounted knights and skeletal dragons for the latest integration.

Tools live in h3-art-pipeline; models and complete mods stay local. Migration and reproduction.

History

Earlier article and revision notes (expand)

The text, images and strikethrough annotations below are preserved from before this editorial pass. “Current”, unfinished work and next steps refer to the time each passage or annotation was written. Old versions, paths and trial renders are historical records.

Status wording from 0.12.0; the current patch version is stated above:

As of September 9, 2026: creature mod 0.12.0 covers all fourteen Necropolis creatures. Measurements here retain their delivery-version scope; see mounted knights and skeletal dragons for the latest integration.

As of September 9, 2026: creature mod 0.11.1 contains ten Necropolis creatures. Installation counts in this article belong to the named delivery version. The latest expansion is Vampire and bat forms.

2026-09-09 tool migration: Development continues in h3-art-pipeline. The old tools/creature-art/ and tools/town-art/ paths correspond to creature-art/ and town-art/ in the new repository. Historical paths and PR links remain in this article. Migration and reproduction details.

Last post ended with eight of fourteen Necropolis concepts passing the proportion check. Six were too wide, none too narrow. This is how that got fixed, and the first 3D model.

Adjectives made it worse

A one-sided error suggests an obvious fix: tell the model to tighten up. I derived a phrase per creature from the measured aspect — the bone dragon, at 0.76, got “broad and low, only slightly taller than it is wide”.

It went from +35% to +51%.

Obvious in hindsight: adjectives describe a pose as much as a shape. “Broad and low” reads like a description of a stance — wings out, body dropped — and the model obliged. The wraith overcorrected the other way, +16% to -18%.

A number doesn’t have that problem:

The whole silhouette fits in a box about 1.3 times as tall as it is wide.

It constrains the bounding box and implies nothing about the pose. It also comes straight from the original’s measured aspect, so it’s one line of code rather than fourteen hand-written hints.

All fourteen now land inside the 15% band. The bone dragon went +51% to +9%.

Fourteen Necropolis concepts
Fourteen green. The previous pass had eight.

The content filter blocks zombies

Two undead prompts were rejected by FLUX as Violence. The triggers were phrases like “rotting flesh”, “dried blood” and “bloated”.

Rewording to desiccated, grey-green skin drawn tight, tattered rags passed, and the creature is unchanged. That’s recorded in a comment in the roster file — left unexplained, someone tidies the wording later and walks straight back into a rejection.

The first mesh

With the concept settled, the skeleton went to Meshy for image-to-3D: quad topology, remeshed to 20k, textured.

The skeleton as a 3D mesh
37508 vertices, 46034 triangles, base colour texture. 30 credits.

The image is cropped to its subject by alpha before upload. A concept sits on a 1024×1440 canvas with the subject occupying 19% of it — reconstruction works on what you hand over, and framing it costs nothing.

The split is deliberate: concepts stay with FLUX.2 [pro] from Black Forest Labs, Meshy does this stage only.

Update, 2026-09-09: The current workflow uses imagegen concepts, Meshy textured meshes and suitable humanoid auto-rigs, with Astra-authored local repair, rigging, motion and packaging. The footless Wight uses a local rig; Tripo is not required. Shadows/outlines are now prebaked offline. See the workflow change and Lich delivery.

Next

Rigging. The skeleton is humanoid so Meshy’s rigging API covers it, but most of the roster isn’t, and that part needs something else.

Code in PR #10: the proportion hint, the reworded prompts, and gen_mesh.py.