InterForge is a local-first AI pipeline for concept artists and game developers. Type a prompt. Get a concept. Refine it to 6 views. Ship the asset. No cloud. No subscription. Your GPU, your files.
Type what you want. SDXL Juggernaut XL generates a batch of concept images tuned to your asset type — character, weapon, prop, tile, environment. Each output comes with a clean RGBA cutout and an SVG silhouette ready for the next stage.
Tuned prompt templates per asset type. Style modifiers from painterly to pixel art.
rembg background removal. vtracer SVG silhouette. Both locked before Smelt opens.
Character, weapon, creature, environment, prop, tile — each with targeted prompt rules.
Painterly, low-poly, pixel art, realistic, stylized — applied per-job via MasterForge.
Lock a concept and Zero123++ synthesizes six camera angles — front, front-right, right, back, left, front-left — at alternating elevations. Consistent silhouette. Consistent lighting. Each view gets its own RGBA cutout and SVG mask.
Multi-view synthesis with 6 fixed camera poses. Alternating ±20° / −10° elevations for clean 3D coverage.
Each of the 6 views: rembg background removal + vtracer SVG silhouette. Sharper masks for reconstruction.
Review each view individually. Approve, reject, or re-run a single angle. Only approved views pass to Forge.
The approved views feed a visual hull carving pipeline. Silhouette masks carve a 3D volume, photo-consistency refines it, and Open3D Poisson reconstruction produces a watertight mesh. The result is decimated into a 4-level LOD chain and exported as GLB, OBJ, or FBX.
SVG silhouettes from all 6 views carve a 3D volume. The intersection gives the base shape.
RGB consistency carves concavities the silhouettes can't catch. Front view weighted highest.
Open3D Poisson reconstruction. Taubin smoothing preserves volume — no shrinkage on organics.
4 decimation levels: 100% / 50% / 25% / 10%. Vertex colors projected from the 6 views. GLB out.
The config engine behind every stage. Asset type templates, style modifiers, lighting presets, and reconstruction routes — all selected per-job. Swap a style or a lighting preset and the prompt chain updates automatically.
Laplacian + Taubin smoothing. Characters, creatures. Preserves volume.
Sharp edge preservation. Weapons, props, architecture.
Disables mesh reconstruction. Sprites, tiles, flat assets.
| Negative Prompts | Auto | per asset type |
| Elevation (Zero123) | ±20° / −10° | 6 fixed poses |
| Smoothing | Taubin | no volume shrink |
| LOD Levels | 4 | 100 / 50 / 25 / 10% |
No cloud, no API keys, no subscription. Every model runs on your hardware. Every output stays on your machine. All job progress streams live over SSE — you see every step in real time.
SDXL and Zero123++ run entirely on your GPU. Nothing leaves your machine.
Every pipeline stage streams real-time SSE events. Step active, step done, image ready — you see it all.
References, sketches, prompts, exports — all saved together per project. Nothing gets lost between sessions.
| Requirement | Spec | Notes |
|---|---|---|
| GPU | RTX 3070 (8GB VRAM) | CUDA 12.1 · target hardware |
| RAM | 16 GB | 8 GB minimum |
| Runtime | Python 3.11 + Tauri 2 | FastAPI backend · React frontend |
| 3D Viewer | Three.js 0.183 | OrbitControls · ACES tone mapping |
| Concept Model | SDXL Juggernaut XL v9 | via diffusers |
| Multi-View Model | Zero123++ v1.2 | 6-view synthesis |
The pipeline enforces stage locks by default — Smelt only opens after Prospect, Forge only after Smelt. Tinker Mode drops the gates entirely. Jump to any stage, re-run any step, experiment without committing. Built for rapid iteration when you know what you're doing.