What I Learned Shipping 30 AI-Generated Game Assets to Roblox in a 48-Hour Game Jam (Using Meshy)

\ I used Meshy and the Meshy Roblox Bridge for asset generation in this jam. This post is about directing a game when your art pipeline isn't the bottleneck anymore. It's not a product review. 1. The Game First, the Tools Second By hour 47, three of us had something we ended up calling Tides of the Drowned. It was a 20-minute Roblox exploration piece across three sunken levels of a civilization the ocean had taken back. You start in the Outer Reef, swim through a Sunken Plaza where broken archways frame a collapsed altar, and end in the Inner Sanctum where a cluster of cyan crystals hums under the last torch the drowned ever lit. There's a trident you can pick up at the end. We didn't have time to give it a use beyond carrying it out. Three people on the team. One programmer, one designer, and me on art. The jam was a weekend on the global circuit. 48 hours, one-word theme: Bubble. We spent the first hour going the other way. What if a civilization ran out of bubbles? That became the game. Tides of the Drowned. Built in Roblox, with all 30 of its 3D models generated by Meshy and pushed into Studio through the Meshy Roblox Bridge. Here's how the 48 hours actually ran. 2. Designing the World Before Writing a Single AI Prompt Before we generated a single asset, the designer and I spent thirty minutes on the world bible. Same way you'd brief a freelance artist if you had one. Five lines: The civilization is gone. The ocean got it. Limestone architecture, weathered, encrusted with cyan coral. Bioluminescent veins ran through everything they built. Some kind of marine symbiote. Palette: muted teal, bone, the faintest cyan glow. Style: stylized low-poly, painterly textures. Not realistic. Then I translated the bible into a prompt template: [OBJECT], drowned civilization aesthetic, weathered limestone with cyan coral encrustation, soft bioluminescent veins, muted teal-and-bone palette, stylized low-poly geometry, painterly texturing, [USE CONTEXT] Every prompt for the next two hours used this exact scaffold. Only [OBJECT] and [USE CONTEXT] changed. The locked modifiers are doing the actual work. "Weathered limestone with cyan coral encrustation" fixes the material. This single phrase is responsible for most of the cross-asset consistency. "Muted teal-and-bone palette" kills the warm-color drift that AI generators tend toward by default. "Bioluminescent veins" is the world's visual signature. It's the one detail every asset carries that signals we belong to the same place. For metal hero props like the trident and the ceremonial mask, I swapped limestone to bronze and kept the rest. The result still read as the same world. Same palette, same encrustation pattern, same level of stylization. One word changed, and the bible held. \ The first batch of eight assets came back between 2,049 and 3,145 triangles. That kept the set comfortably below Roblox's general individual mesh ceiling of 20,000 triangles. Style consistency, eyeballed across the set, was around 80%. The rest of the 30 followed the same template. This is the part I want to land hard. The template wasn't a prompt engineering trick. It was the art bible compressed into a format an AI 3D model generator could execute. Every modifier I locked was an art direction decision I'd already made. The generator wasn't deciding anything. It was implementing. 3. Building Three Levels in Roblox with Meshy Assets Level 1: The Outer Reef The opener. Players swim in from the edge, past the broken pillars and overgrown urns that mark where the civilization once met the sea. The brazier still has its bowl. Everything else is half-claimed by the reef. Six assets here. Broken pillars at 2,049 triangles. Ceremonial urns at around 3,000. The stone brazier at 3,114. Dense coral overgrowth pieces at 2,727 for set dressing. \ The pipeline from Meshy to Roblox is genuinely one click. The Meshy Roblox Bridge is a small desktop app for macOS and Windows that runs a local server on port 5330. You sign into your Roblox account once. From then on, every finished asset in Meshy has a Send to Roblox button. Click it, and the asset lands as GLB in your Creator Hub inventory within a few seconds. Open Roblox Studio, go to Toolbox, Inventory, Meshes, drag the asset into your scene. No download, no FBX juggling, no manual Import 3D dialog. If you're starting from scratch on Roblox 3D imports, Meshy's Roblox modeling guide covers the format and triangle constraints in detail. Worth a skim before your first batch. \n Level 2: The Sunken Plaza The middle act, and the one I wanted players to feel. They round a coral wall and the plaza opens up. A collapsed archway frames a ruined altar at the center, broken pillars stagger out from the perimeter. Scale needs a manual pass. Meshy generates each asset against its own bounding box, so different objects arrive at very different stud sizes. In my batch, the trident's volume reading was 19,416 cm³ while the archway's was 1,395,337 cm³. That didn't translate cleanly into Roblox scale, but it was enough warning that the assets weren't arriving with a shared sense of size. Dropped into Studio without adjustment, the trident read far too small next to the architecture. The fix is fast. Drag the asset in, eyeball it against the player character, scale. Five seconds per asset. There's no batch tool. You do them one at a time. CollisionFidelity matters once the player walks through your geometry. Roblox's Default setting isn't wrong, but it isn't a promise that openings behave like openings. It uses a voxel-based convex-hull decomposition, which is fast enough for many props but can be inaccurate for holes, doorways, and cavities. For a solid pillar, cheaper collision is fine. For the archway, which is supposed to be walkable, it wasn't. The first test made the problem obvious. The player could see the doorway and still bounce off an invisible wall. \n Level 3: The Inner Sanctum The end. Cyan crystal shards cluster around a centerpiece altar. The ceremonial mask is mounted into the wall. The trident sits on a plinth in the back. This is the climax, and visually it's the densest scene in the game. Meshy generates PBR maps alongside the geometry. Base color, normal, metallic, roughness. Roblox's SurfaceAppearance object has four corresponding slots: ColorMap, NormalMap, MetalnessMap, RoughnessMap. Uploading each texture into the matching slot gave me the closest starting point. I still had to judge the result inside the actual Roblox lighting setup. The crystals were semi-transparent, and this is where I chose the jam-speed version of the material workflow. I used SurfaceAppearance for the PBR maps, then used the MeshPart Transparency property as a fast global art-direction slider against the Sanctum lighting. Roblox can also handle transparency through SurfaceAppearance.ColorMap alpha and AlphaMode, but under a 48-hour clock I wanted one control I could tune visually in context. The crystals end up reading as luminous in the Sanctum's low ambient light, which is the effect the level needed. Thirty assets, three levels, one consistent visual language. I spent four hours on art across all of it. 4. The 20% of AI Assets That Needed Another Pass Across 30 generated assets, some didn't quite fit the world on the first pass. Color drift. A few came back warmer than the palette called for. One trident variant landed brassy-orange instead of weathered cyan-bronze. Detail density mismatch. A simple ceremonial urn that came back with intricate filigree the rest of the world didn't have. Scale anomaly. One asset that should have been a centerpiece statue arrived hat-sized. Meshy's bounding box had decided the prompt meant small. This is the part of the workflow that's least talked about, and the part I want to be direct about. First-pass AI assets don't all land. Six out of thirty in this run, which lines up with what experienced art directors will tell you about any first-pass review. AI or human. The thing that matters isn't whether you get a 100% hit rate. It's how cheap the iteration loop is. Two paths I used. First path. Refine in Roblox. Keep the geometry, adjust SurfaceAppearance, swap in a tinted ColorMap, or use SurfaceAppearance.Color for a quick tint pass. About 30 seconds per asset. This covers color and tone fixes. If the geometry is wrong, no amount of tinting helps. Second path. Regenerate in Meshy. Same prompt, hit generate again. Text-to-3D is non-deterministic. The same prompt produces different results on different runs, and sometimes the next run is exactly the version you wanted. About 90 seconds of generation plus another Bridge push. This covers geometry and scale issues. For context, modeling six replacement assets from scratch in Blender would have cost much more time. The iteration cost on AI generation is genuinely an order of magnitude lower than the iteration cost on traditional modeling. That's the part of the workflow that matters, not the first-pass hit rate. 5. Where the Twelve Hours Went The cleanest way to describe the time split isn't total person-hours, because we overlapped constantly. This is the critical-path shape of the 48 hours: Art (generation, Bridge pushes, Studio import, rescale, refinement): 4 hours Level design: 20 hours Programming: 20 hours Testing and final build: 4 hours Twelve hours that used to go into making assets exist went instead into making them mean something. That's the moment in Level 2 where the player rounds the coral wall and the plaza opens up. That beat needs 20 hours of pacing work to land, not 8. It's the lighting falloff in the Sanctum that lets the crystals carry the scene without competing with the centerpiece. It's the placement of the trident on its plinth, angled so the player sees it from across the level before they cross to reach it. None of that is rendering. All of it is direction. We shipped 30 unique assets, three playable levels, and a complete internal demo by hour 47. 6. What AI 3D Generation Changes About Game Art Direction Before this jam I had two assumptions about generative AI for game art. Both turned out to be wrong. The first assumption was that AI tools would either replace what I do or be useless to me. Neither happened. What actually happened is the four hours I spent on art looked completely different than they used to. Not less work, different work. Less time spent making assets exist, more time spent deciding what assets the world needs, what they should communicate, which ones earn their place. The second assumption was that you couldn't get art direction out of a generator. That consistency across a large asset set was something only a human artist could hold. That turned out to be wrong in a specific way. The consistency doesn't come from the generator. It comes from the bible. The template I built in section 2 was just an art bible compressed into a format the generator could execute. The world held together because the bible held together. The generator implemented. The direction was mine. If you're a solo dev or a small team thinking about trying AI generation for a Roblox project , one specific piece of advice. Write the bible first. Then write the template. Then generate. Don't open the generator until you can answer, in one sentence, what your world is made of and what it isn't. The teams that get inconsistent AI output are the teams that started typing prompts before they'd done the direction work. 7. FAQ Q1: How does the Meshy Roblox Bridge work, and what platforms does it support? A desktop app for macOS and Windows. It runs a local server on port 5330. Sign into your Roblox account once. Every finished Meshy asset then has a Send to Roblox button. Click it and the asset lands as GLB in your Creator Hub inventory within seconds. In Studio, go to Toolbox, Inventory, Meshes, drag it into your scene. Q2: Do Meshy's exported textures work directly in Roblox? Mostly. Meshy generates four PBR maps alongside the geometry: base color, normal, metallic, roughness. These map one-to-one to Roblox SurfaceAppearance's four slots: ColorMap, NormalMap, MetalnessMap, RoughnessMap. Upload each into the matching slot for a starting point. You'll still want to tune the result inside your actual scene, since PBR appearance depends on the lighting setup and the user's device, not just the maps. Q3: How do you maintain visual consistency across many AI-generated assets? Write a locked prompt template before generating anything. In our run, five modifiers never changed across the asset set: material, palette, signature detail, geometry style, texture style. Only the object name and use context varied per prompt. The consistency comes from the template, not the generator. Q4: Is this workflow only useful for game jams, or does it work for production-scale projects? It maps most cleanly to jam, prototype, and vertical-slice work, where art is the bottleneck between idea and shippable build. For production-scale projects, AI generation is closer to a first-pass concept tool. The direction work doesn't go away. It gets more time. Meshy's guide to AI tools for Roblox game development covers the production-side workflow in more detail. Q5: Does AI generation replace the artist on the team? No. It changes what the artist's hours look like. Less time spent making assets exist, more time spent deciding what the world needs and what each asset should communicate. The generator implements. The direction is still yours.