--- title: Generating Game Assets with AI Agents description: The field-tested playbook for making image→3D models, textures, and audio for games on Wireflow — including the failures, so your agent gets it right faster. updated: 2026-10-14 --- This guide is the distilled playbook from building [Breach Point](https://breach-point-cmrou18ll0.wireflow.ai), a multiplayer arena FPS where **every asset — wall textures, the rifle, the enemy drone, the first-person hands, the sound effects — is an upstream Wireflow node**. It took eight iteration rounds to get the viewmodel right; each failure below is one your agent can skip. Everything here was produced by an agent driving the public API, so every rule is agent-actionable. The full workflow is public: [inspect the asset graph](https://www.wireflow.ai/workflows/cmrou18ll0001xgk34fdgvyyj). ## The pipeline Game assets are a two-stage chain, never a single step: ``` input:text → image model (concept art) → image-to-3D model (GLB mesh) → game node port input:text → audio model (SFX) → game node port ``` The image model invents the look; the 3D model reconstructs geometry from it. **What the concept image shows is what the mesh becomes** — which is why prompt craft is 80% of the outcome. ## The 9 rules that matter ### 1. The concept image IS the mesh's camera An image→3D model reconstructs the exact pose it sees. If your concept is a "three-quarter product shot", your mesh is a product shot, and it will look wrong in an FPS camera forever. Compose the concept **like an in-game screenshot**: gun at bottom-right, muzzle toward screen center, plain background. ### 2. Weapon and hands as separate meshes The industry standard, and the single highest-leverage decision. A combined hands+rifle mesh reads beautifully in concept art but its in-engine alignment is a lottery (see rule 7). Keep a clean weapon GLB and a separate hands GLB — you can always swap or re-tune one without the other. ### 3. One material family per mesh Dark gun + dark gloves + dark sleeves reconstructs as one unreadable blob, for both the image model and the 3D reconstruction. Give each mesh a tight palette with deliberate contrast (dark gunmetal + warm orange accents works). ### 4. Thick, simple organic shapes; sleeves over skin Long thin parts melt first — forearms, fingers, antennae. Full sleeves beat bare skin: one less material, no fingers, and sleeves come out as clean tubes. Our first hands mesh had melted, stubby forearms; cropping them into sleeves fixed it. Note that artifacts at **crop boundaries** (where the concept cuts the object off, like a sleeve cuff) are seed-dependent — if the mesh has a wispy shard there, a same-concept re-roll often fixes it for $0.10 before you touch the prompt. ![Melted forearms on a generated viewmodel](/images/docs/fail-melted-forearms.jpg) ### 5. No contradictory prompts "hands **and arms**" plus "**no forearms visible**" confuses the image model — it picks one. Short, focused sentences; say what you want, not what you don't. The prompt that failed: ``` …two tactical gloved hands and arms gripping a rifle… short grey tactical cuffs at the wrist only, no forearms visible… ``` ### 6. Prompt for reconstruction, not beauty What survives image→3D: bold silhouettes, uniform materials, even lighting, plain backgrounds. What dies: dramatic shadows, depth of field, fine engravings, skinny details. Write "game asset, studio product lighting, plain solid light grey background" and mean it. ### 7. Judge the mesh from the exact in-game camera — and center by geometry, not bbox A mesh normalized by its bounding box centers the _arms_, not the gun. Anchor by measured geometry (e.g. compute the barrel tip from the forward-most vertex slice) and render-candidate framings against your real camera FOV/offset before shipping. Per-asset viewmodel offsets (position/euler/scale) are completely normal — every FPS game hand-tunes these per weapon. **Judge candidates against a reference frame from a real game, not against your last attempt** — a framing that merely beats the previous one can still be wrong (we shipped a muzzle-points-at-the-floor hip pose that way). ### 8. Expect per-asset orientation constants 3D generators emit arbitrary poses. Auto-orient heuristics (e.g. "the thin end is the muzzle") work for clean hard-surface meshes and **fail on combined organic+hard meshes** — our wrist-cuffed mesh shipped backward because the thin wrist stubs read as "the muzzle end". The fix: render a turntable of every GLB, then pin a verified yaw constant per asset. ![The same mesh shipped backward — always turntable-check orientation](/images/docs/fail-backward-gun.jpg) ### 9. Audio: short transients + pitch jitter, or don't bother A 0.8s gunshot sample retriggered every 0.13s stacks ~6 tails into echo-mush. Trim shots to a ~0.3s transient (analyze attack time and crest factor — you don't need to listen), normalize under 0dBFS, and add ±5% `playbackRate` jitter per shot so full-auto doesn't sound like a loop. ## Picking the image→3D model (July 2026) | Model | Wireflow node | Cost | Multi-view | Notes | | ---------------- | --------------------- | ------- | ------------------------------------------------- | ----------------------------------------------------------------------------------------------------- | | Trellis 2 | `generate:trellis_2` | lowest | via API (`image_urls`), not yet exposed as a port | Cheapest; great hard-surface. Our default for props. | | Hyper3D Rodin v2 | `3d:hyper3d_rodin_v2` | high | **Yes** (`input_image_urls` port) | Multi-view conditioning today — the winner for organic+hard combined meshes (arms+weapon viewmodels). | | Tripo H3.1 | `3d:tripo_h31` | mid | No | PBR + texture/geometry quality knobs. | | Meshy 6 | `3d:meshy_v6` | highest | No | PBR, polycount control, rigging options. | Multi-view (3 angles of the same object) improves **fidelity**, not framing — it's the fix for "the back of the mesh is guesswork", not for "the pose is wrong". But note it _is_ the decisive lever for organic+hard combined meshes: our hands+rifle viewmodel failed 4× in single-view (melted forearms, crop shards) and passed first try in Rodin multi-view. For web games, set `decimation_target` 20k–50k and `texture_size` 1024–2048; defaults (500k) produce multi-MB meshes — and even then, budget a gltf-transform pass (texture resize + simplify) to ship under ~5MB. ## The verification loop (cheap, catches everything above) 1. **Dry-run the graph** ($0): `npx tsx scripts/dryrun.ts ` — routing/schema/port errors before a cent is spent. 2. **Preflight credits**: `npx tsx scripts/preflight-credits.ts ` — the balance must cover estimate × 3. 3. **Render the GLB offline** before shipping: an 8-angle turntable + a render from the _exact_ in-game camera. This caught a backward-facing enemy drone and a backward rifle before players saw them. ![Turntable review of the generated enemy drone — the eye faces +Z, which the code had wrong by 90°](/images/docs/enemy-turntable.jpg) 4. **Then** a live screenshot of the real game. ## What the end state looks like ![Final viewmodel: full rifle angled into the crosshair, both hands on the gun](/images/docs/final-viewmodel.jpg) ## Copy-paste block for your agent ```text You are generating game assets via Wireflow workflows. Rules: 1. Assets are two-stage chains: input:text → image model (concept) → image-to-3D (GLB). Never call a model API directly. Build a scratch workflow, dry-run + preflight it, then POST /api/v1/workflows//run. 2. Concept prompts: short sentences, one material family per mesh, plain light-grey background, "game asset, studio product lighting". Compose viewmodels like an in-game screenshot, not a product shot. 3. Weapon and hands as SEPARATE meshes. Sleeves over skin. No contradictory instructions in prompts. 4. Trellis 2: decimation_target 30k-50k, texture_size 2048 for hero assets, 1024 otherwise. 5. After generating: render an 8-angle turntable of the GLB, pin per-asset yaw/offset constants, verify from the exact in-game camera, THEN ship. 6. Audio: <=0.35s transients for retriggered sounds, normalize to -3dB. ``` ## Cost reality A full asset suite for a small game (4 textures, 2 sprites, 3 GLB models, 3 SFX) is cheap, iteration rounds included. Sound effects and concept images are the cheap steps and 3D meshes the expensive one, so the two-stage split pays: a bad concept is cheap to throw away, a bad mesh is not. Last updated: 2026-10-14 --- Documentation index: fetch https://www.wireflow.ai/llms.txt for the full list of Wireflow docs.