Since my last post last post I’ve been playtesting a lot and improving the gameplay, but have been a bit stuck on how to create good AI assets.  I have the 5090 and a pile of local image models.  My initial attempts at making a basic set of buildings were kinda lackluster.  It worked, but the images were either too cartoony, or had too much added stuff, or just looked weird.  I ended up biting the bullet and buying Meshy rather than continuing to tweak local models and eventually settled on a pipeline that works.  Relying on Claude to generate image prompts worked surprisingly well, but I still need to do quite a bit of manual intervention to make the models work.

Pipeline

Here’s the pipeline that’s been working for me.  I start by generating the images and tweaking there, then feed into Meshy to make models themselves

Concept PNG  →  Mesh only  →  Blender cleanup  →  Texture  →  Godot
   (Flux)         (Meshy)      (strip base)       (Meshy)    (live render)

1. Concept image (local Flux) — ComfyUI on the Windows side with Flux.1 Dev and a building-biased LoRA.  Each building has its own prompt file plus a shared style sheet: industrial colony massing, three-quarter view, plain studio background, and a long list of don’ts (no glass domes, no toy aesthetic, no pedestal).  If the silhouette is wrong here, nothing downstream saves it.

2. Mesh only (Meshy) — Image-to-3D with texturing turned off.  Remesh toward a medium-poly budget (~4k–8k faces).  Texture comes later on purpose so cleanup doesn’t fight baked materials.

3. Blender cleanup — AI almost always bakes a flat display base under the building.  I manually strip that pad, set origin to ground-center, and clean any other weird artifacts.  The mesh is then checked against a Blender authoring template so scale, front axis, and footprint match the game’s rules before anything is textured.

4. Texture (Meshy again) — PBR on the cleaned mesh only.  I reapply the original image so the texture style doesn’t drift.

5. Godot — Approved bld_*.glb files drop into the project.  The engine adds a terrain-conforming pad at placement time and the model plugs into the same slot in the code as the placeholder boxes.

Animation is intentionally not Meshy character animation — conveyor belts, rotors, and smoke get driven in-engine once the static mesh is solid.  I’ve only done a couple of these as a proof of concept, but they’re pretty simple animations so shouldn’t be too bad.

Where things stand

I’ve generated some basic buildings that I’m happy with and have hand-placed animations for them.

     

Now I just need to keep cranking them out and then figure how to handle zoned buildings.  These will definitely require the largest volume of assets and I’m thinking about generating some building block assets first and then combining them in various permutations for each building size.  I’ll report back once things are polished up, but don’t really want to release the game until it’s polished.  I might even try to publish it to Steam but I’d want to polish it even more for than than I would for even a beta demo.