Running AI image generation locally usually means wrestling with Python environments, CUDA drivers, and ComfyUI node graphs. Mold takes a different approach: it's a single Rust binary that handles text-to-image, image-to-image, and inpainting from your terminal.
Built on the Candle framework (a pure-Rust ML library), Mold requires no Python, no libtorch, and no cloud connection. You install it with a one-line curl command, point it at a model, and generate images.
What it supports:
- 8 model families, 60+ model variants: FLUX.1 (schnell, dev, plus fine-tunes), Flux.2 Klein, SDXL, Stable Diffusion 1.5 and 3.5, Z-Image, Qwen-Image, and Wuerstchen v2
- LoRA adapters (custom style models) in both BF16 and GGUF quantized formats
- ControlNet support for guided generation (canny edge detection, depth maps, pose estimation)
- Local LLM prompt expansion using Qwen3-1.7B to improve your text prompts automatically
- Embedded metadata in output PNGs so you can reproduce any image from its seed and settings
- A REST API with streaming, auth, and rate limiting for building apps on top
- A Discord bot with slash commands
Models download automatically on first use (2-30GB depending on the model). You need an NVIDIA GPU with CUDA or Apple Silicon with Metal support.
Mold is MIT-licensed, written almost entirely in Rust (97.5%), and currently at v0.4.1 with 172 commits across 7 releases. It's very early - 6 GitHub stars as of today - but the feature set is ambitious for a CLI tool. The Unix pipe compatibility is a nice touch for scripting workflows, and terminal preview rendering (via Kitty, Sixel, or iTerm2 protocols) means you can see results without leaving your terminal.
For anyone already comfortable at the command line who wants local image generation without the Python dependency overhead, this fills a real gap. The tradeoff is that you lose the visual node-graph workflows of tools like ComfyUI, so it's better suited for batch generation or integration into automated pipelines than for iterative creative exploration.