ewk labs / forge / about

Technical Architecture

STEP 00

Train

LoRA fine-tuning from reference dataset

FLUX.2 Klein 4B

STEP 01

Generate

Text-to-image with optional trained style

FLUX.2 Klein 4B

STEP 02

Multi-View

Front, back, top, bottom synthesis

FLUX.2 Klein / Zero123++

STEP 03

3D Model

Multi-view mesh reconstruction

Hunyuan3D 2.1

STEP 04

Scale, repair, flatten, export STL

Trimesh + NumPy

Models

FLUX.2 Klein 4B Steps 00, 01, 02 / Training + Generation

Black Forest Labs' fastest model combining text-to-image generation with multi-reference editing in a single 4B parameter architecture. Sub-second inference with quality that rivals models 10x its size. Powers LoRA training (Step 0), initial image generation (Step 1), and text-guided multi-view synthesis (Step 2).

Parameters 4B (rectified flow transformer) Inference 4 steps (step-distilled) Training DreamBooth LoRA, 500-1000 steps on A100 VRAM ~13GB (inference) / ~40GB (training) License Apache 2.0 By Black Forest Labs

Zero123++ Step 02 (alt) / Geometric Multi-View

Generates 6 geometrically consistent views from a single reference image using learned 3D priors. Unlike FLUX's text-guided approach, Zero123++ produces views with strict camera angle consistency, making it better suited for 3D reconstruction. Includes automatic background removal via rembg.

Output 6 views (front, back, left, right, top, bottom) Method Diffusion with 3D-aware conditioning Post-process rembg background removal + ONNX Runtime By SUDO AI

Hunyuan3D 2.1 Step 03 / 3D Reconstruction

Tencent's large-scale 3D generation system. Converts multi-view images into detailed 3D meshes using a flow-based diffusion transformer for shape generation. Configurable octree resolution and guidance scale for balancing quality vs. speed. Outputs GLB format with PBR-ready geometry.

Shape Model 3.0B parameters Output GLB mesh (downloadable) Resolution Configurable octree (256-768) Best Settings Steps 75-100, Guidance 10-12, Octree 512 By Tencent Hunyuan

Post-Processing

Print-Ready Pipeline Step 04 / Scale + Repair + Export

Prepares 3D models for physical fabrication. Scales mesh to target shoe size (US 8-13), flattens the sole for proper contact surface, fills holes, repairs normals, and exports watertight STL for 3D printing. Runs on CPU-only infrastructure.

Stack Trimesh + NumPy + fast_simplification Input GLB from Step 03 Output STL (print-ready, watertight) Sizing US 8-13 (25.7-31.8 cm)

Infrastructure

Modal Serverless GPU Compute

All models run on-demand on NVIDIA GPUs with automatic scaling. Pre-baked container images with cached weights for fast cold starts. Per-second billing with no idle charges. 6 active endpoints across the pipeline.

GPUs NVIDIA A10G (24GB) for inference, A100 (40GB) for training Endpoints 6 (FLUX, Hunyuan3D, Zero123++, Training, PostProcess, Vision) URL modal.com

Three.js 3D Viewer

Browser-based 3D rendering with orbit controls, studio lighting (key + fill + rim), and PBR material support. Loads GLB models directly for interactive inspection. Used in both Step 3 (reconstruction preview) and Step 4 (print-ready preview).

Version r160 (ES Modules) Features OrbitControls, GLTFLoader, ACES tonemapping, auto-rotate

GitHub Pages Frontend Hosting

Static frontend served from GitHub Pages. Zero-config deployment via git push. All compute happens server-side on Modal; the frontend is pure HTML/CSS/JS with no build step.

Domain forge.ewklabs.xyz Stack Vanilla HTML/CSS/JS, no framework