[AI] Improving Heroes III with AI

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.

Update, 2026-09-09： The table preserves the original tool choices. Struck-through entries were superseded; Meshy remains in use for base geometry and some humanoid rigs.

Heroes III came out in 1999. Twenty-seven years later people still play it, and a group of people are still rewriting its engine from scratch.

This series documents a hobby project: standing on VCMI and pushing the whole game forward. For fun, not commercial.

The foundation

VCMI is an open-source Heroes III engine, reimplemented from scratch and compatible with the original’s gameplay and assets. It’s a good base not because it runs, but because the extension points are already there: per-scale HD sprite trees, JSON-defined animations, data-driven creatures and skills, Lua wired into combat.

So the tracks below can move independently.

The workflow

What is used where, and why:

The split is deliberate: concepts stay with FLUX, Meshy does mesh and rig only, Blender does animation and render.

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.

Every stage writes a json alongside its output recording the task id, parameters and cost, so a result traces back to the request that produced it.

Art

The whole creature roster replaced, through a 3D pipeline: concept → image-to-3D → rig → animate → render back to sprites.

One counterintuitive detail: this isn’t a betrayal of the pixel look, because Heroes III’s creatures were pre-rendered from 3D to begin with. The skeleton’s palette holds 223 distinct colours — never limited-palette pixel art, just a render squeezed into a low resolution. Going 3D is a return to the original method.

Under two dollars per creature, roughly $280 for all 150. The money isn’t the problem. Art direction and per-creature review are what take time.

A stronger AI

This one already works — it isn’t a proposal.

smanolloff wrapped VCMI battles as a gym environment and trained models on it. The result, MMAI, has been selectable in the launcher since 1.7.0 and lives in AI/MMAI/ in the main tree: ONNX Runtime inference, separate models for attack, defence and siege, temperature sampling, fallback to the classic AI if inference fails.

The original combat AI is decades-old heuristics — predictable and exploitable. So this track starts from working code rather than from scratch.

Mechanics

The engine is data-driven, so changes are cheaper than they sound: terrain effects, stack synergies, adventure-map pacing.

A side effect may matter more than any of it: using the RL environment as a balance tool. Change one number, self-play tens of thousands of battles, watch the win rate move. Balance testing with objective, fast feedback.

The bad idea: non-blocking turns

The worst part of turn-based multiplayer is waiting. The existing mitigation is simultaneous turns — everyone moves at once until two players meet, then it collapses back to alternation.

I want to try fully asynchronous: everyone moves at their own pace, conflicts resolved only where interaction actually happens. The difficulty is entirely there — two players grab the same mine at their own “current moment”, who got there first? Introduce a real timeline and is it still turn-based?

I haven’t worked either of those out.

On tooling and rights

Coding agents got good enough that we dared to start at all. The code is written with Claude Code, the concept art with FLUX.2 [pro] from Black Forest Labs. Reading 100k lines of unfamiliar C++, reverse-engineering the .def format, writing a validator — weeks of prerequisite work, now a day or two.

The tooling lives in tools/creature-art/ in yzh119/vcmi.

Heroes III’s assets are Ubisoft’s, so derived assets aren’t distributed; what we build is tooling and original art. VCMI is a decade of other people’s work — anything useful goes back upstream.

Written so far

• the engine and the archive formats — the rules that never error, and how three images hide inside one palette
• two metrics I got wrong — how the original art caught both

Next: mesh, rig, animation, rendering back to sprites.

Update, 2026-09-09: Eight Necropolis units are installed in local0.10.0. Vampire and Vampire Lord are in progress. See the latest delivery.

First target: one skeleton, end to end.