The Idea

I wanted to build a small atmospheric scene for the devlog, a Welsh street at night, terraced cottages, a fox with a scarf walking past a gas lamppost. The kind of thing that takes a professional 3D artist a few days to set up properly.

I know almost nothing about Blender. I've poked around in it, moved a few things, broken a few more things. So this was partly an experiment in how far you can get with AI doing the heavy lifting while you steer the creative direction.

Planning with Meshy

Rather than jumping straight into making assets, I used the Meshy AI agent to think through what the scene actually needed. I described the mood, the setting (terraced cottages, narrow pavements, iron railings, window boxes), and what story the scene should tell.

The agent helped me break this down into a modular asset list:

Welsh terraced cottages (a few variants)
Street lamppost cast iron, Victorian style
Pavement slabs and kerbing
Garden walls with iron railings
Hill slope terrain tiles
A fox character in a Welsh scarf

The key thing the agent pushed me towards was thinking modularly, not one big scene mesh, but individual pieces I could place and reuse. This ended up being really valuable.

Generating the Assets

With the asset list defined, I generated each piece through Meshy. The prompts were fairly descriptive — specifying the Welsh aesthetic, the materials (sandstone, red brick, dark iron), and the scale each piece needed to work at in a modular grid.

Meshy's AI model handled the geometry and PBR textures (albedo, roughness, metallic, normal maps) all in one pass. Some pieces needed a second attempt with tightened prompts, but most came out usable first try.

The fox character in particular came out better than expected, the Welsh scarf detail reads clearly even at game scale.

Importing into Blender via the Bridge

This is where it used to get painful. In the past, getting Meshy assets into Blender meant exporting GLBs, dealing with material import issues, fixing texture paths, and a lot of manual cleanup.

The Meshy Blender bridge changed this completely. Assets import directly with all their materials intact — PBR nodes wired up, textures packed into the blend file, ready to place.

Within about twenty minutes I had all the cottage pieces, the lamppost, pavement tiles, and the hillside terrain imported and sitting in the scene. The modular approach paid off here — I could duplicate the cottage mesh and offset it to create the terraced row, then arrange the pavement slabs along the street.

Scene Composition

Putting the scene together was honestly more like playing than working. Drop the cottages in. Lay down the pavement. Place the iron railings along the garden wall. Position the fox by the left lamppost.

I also added a Welsh hillside behind the cottages procedurally generated using Geometry Nodes with a Y-axis slope that rises away from the street. The `HillsideMat` material uses a Voronoi + Noise texture blend to get the vivid, slightly stylised Welsh grass colour.

The result: a compact diorama about 12 metres wide that reads as a real Welsh street from camera distance.

Lighting with the MCP Blender Plugin

This is where it got interesting and where I learned the most.

I have zero experience setting up Cycles lighting in Blender. Exposure, energy units, DOF, world background nodes, all of it was foreign to me. So I used the MCP Blender server plugin, which lets an AI assistant communicate directly with the live Blender session via its Python API.

I described what I wanted: "Night scene. Lamppost as the key light. Moonlight feel, soft blue in the shadows."

1. Placed a warm amber Point light (350W) at the lamppost globe position to cast a cone downward onto the fox and pavement

2. Added a cool blue Sun light at 0.25W from a high angle mimicking directional moonlight, creating subtle blue-tinted shadows on surfaces facing the sky

3. Set the world background to a solid deep midnight navy instead of the evening sky texture (which was previously washing everything orange)

4. Calculated the exact camera-to-fox distance and set the Depth of Field focus precisely on the fox head, f/3.5 fox sharp, cottages and lamppost softly blurred behind

It explained each step as it went, why a Sun light for moonlight (it's directional), why the world background matters for filling shadows with blue, what aperture f-stop does to the focal plane. I understood more about Blender lighting in this one session than I had from watching tutorials.

The Results

I'm genuinely happy with how this turned out. The warm amber pool under the lamppost, the dark Welsh terraces with their brick arches and window boxes, the fox mid-stride on the pavement.

There was friction along the way. Blender's Metal GPU doesn't support rendering in background/headless mode on macOS, so all the CLI renders had to run on CPU (about 2–3 minutes per frame at 256 samples). The compositor node group from a previous session was silently intercepting all renders and outputting black that took a while to diagnose. Meshy's packed textures had lost their file extension metadata, causing everything to fall back to white until we restored the `.jpg` headers.

None of this would have been quick to debug without the AI walking through it step by step.

Takeaways

What AI-assisted 3D actually looks like in practice:

The creative direction is yours. The AI can generate assets and wire up nodes but it doesn't know you want the scarf to be red and white, or that the lamppost needs to be dark iron not brass.

Modular planning upfront saves hours. The Meshy agent pushing me to think in discrete reusable pieces rather than one monolithic mesh paid dividends throughout.

The MCP bridge between the AI and Blender is genuinely powerful. Being able to say "the light should feel like moonlight" and have it translate that into actual node values, explain the reasoning, and let you preview and iterate, that's a different workflow from anything I've used before.

You still have to engage with the tools. I ended up learning more about Blender's lighting system, compositor, and DOF than I expected, precisely because the AI was explaining what it was doing at each step rather than just doing it silently.

I went into this not knowing much about Blender. I came out of it still not being a Blender artist, though I really think understand it a tad better, and I have a scene I can work on.

Seren - Devlog 2 - Blender