SceneWorks

SceneWorks is an open-source Rust framework designed for building scenario-based AI applications. It provides a modular, high-performance runtime for orchestrating AI agents, image generation workers, and inference pipelines — all built with a focus on speed, reliability, and GPU acceleration.

Hosted on GitHub at github.com/SceneWorks/SceneWorks, the project has accumulated over 3,100 commits and 23 branches, reflecting active development. SceneWorks leverages Rust's memory safety and performance characteristics to deliver a robust platform for AI workloads that require low latency and high throughput.

• Rust-Native Runtime - Built entirely in Rust for maximum performance, safety, and memory efficiency.
• GPU-Accelerated Workers - Rust GPU image processing workers using Candle framework for model inference.
• Modular Crate Architecture - Organized into reusable Rust crates for data, design, workers, and apps.
• Image Generation Pipelines - Supports multiple image backends including Krea 2 Turbo and Anima with quantization.
• Design System - Built-in design tooling for generating and processing creative assets.
• Docker & CI - Comprehensive CI pipeline and Docker deployment support.

SceneWorks is organized as a Rust workspace with several key crates under crates/, each handling a specific domain of the system. The framework uses a worker-based architecture where different AI tasks (image generation, data processing, inference) are handled by dedicated, independently deployable workers that communicate through a shared runtime.

The apps/ directory contains application binaries that tie together workers and pipelines. The config/manifests/ directory holds Kubernetes and Docker deployment configurations. SceneWorks uses mlx-gen for Apple Silicon acceleration and Candle for cross-platform GPU inference, with built-in quantization support (q4, q8, bf16) for efficient model deployment.

A distinguishing feature of SceneWorks is its Rust GPU worker infrastructure. These workers handle model inference directly in Rust, avoiding the Python overhead common in AI pipelines. The Rust GPU workers support:

• Candle Inference - Cross-platform ML inference with support for Krea 2 Turbo img2img models.
• Anima Integration - Install-time model conversion with q4/q8/bf16 quantization support.
• Image Processing - Native Rust image conversion pipeline with true_v2 format support.
• Router System - Intelligent routing of inference requests across available GPU workers.

The sc-3019 milestone introduced the Rust GPU image worker scaffold and mlx-gen link, enabling Apple Silicon GPU acceleration directly from Rust.

SceneWorks includes a comprehensive image processing pipeline capable of handling multiple image generation and transformation workflows. The image pipeline supports text-to-image and image-to-image generation, format conversion, and asset management. The sc-3032 update completed the image cutover to a native Rust true_v2 converter, retiring legacy Python-based conversion tools.

The framework includes a design workspace under design/ that provides asset management, style definitions, and creative workflow tooling for building scenario-based visual content.

SceneWorks includes a built-in design system (design/ crate) that provides tooling for creative asset generation and management. This enables the framework to not only orchestrate AI inference but also handle the full lifecycle of creative content — from prompt engineering and style definition to asset generation, processing, and storage. The design system is deeply integrated with the worker architecture, allowing designers to define visual scenarios that the GPU workers then render.