mold

Generate images from text on your own GPU. No cloud, no Python, no fuss.

Documentation | Getting Started | Models | API

mold run "a cat riding a motorcycle through neon-lit streets"

That's it. Mold auto-downloads the model on first run and saves the image to your current directory.

Install

curl -fsSL https://raw.githubusercontent.com/utensils/mold/main/install.sh | sh

This downloads the latest pre-built binary to ~/.local/bin/mold. On Linux, the installer auto-detects your NVIDIA GPU and picks the right binary (RTX 40-series or RTX 50-series). macOS builds include Metal support.

Nix

nix run github:utensils/mold -- run "a cat"                   # Ada / RTX 40-series
nix run github:utensils/mold#mold-sm120 -- run "a cat"        # Blackwell / RTX 50-series

From source

cargo build --release -p mold-ai --features cuda    # Linux (NVIDIA)
cargo build --release -p mold-ai --features metal   # macOS (Apple Silicon)

Add preview, expand, discord, or tui to the features list as needed.

Manual download

Pre-built binaries on the releases page.

Usage

mold run "a sunset over mountains"                    # Generate with default model
mold run flux-dev:q4 "a turtle in the desert"         # Pick a model
mold run "a portrait" --width 768 --height 1024       # Custom size
mold run "a sunset" --batch 4 --seed 42               # Batch with reproducible seeds
mold run "oil painting" --image photo.png              # img2img
mold run "a cat" --expand                              # LLM prompt expansion
mold run flux-dev:bf16 "portrait" --lora style.safetensors  # LoRA adapter

Inline preview

Display generated images directly in the terminal (requires preview feature):

mold run "a cat" --preview

Piping

mold run "neon cityscape" | viu -                     # Pipe to image viewer
echo "a cat" | mold run flux-schnell                  # Pipe prompt from stdin

Terminal UI (beta)

mold tui

Model management

mold list                    # See what you have
mold pull flux-dev:q4        # Download a model
mold rm dreamshaper-v8       # Remove a model
mold stats                   # Disk usage overview
mold clean                   # Clean orphaned files (dry-run)
mold clean --force           # Actually delete

Remote rendering

# On your GPU server
mold serve

# From your laptop
MOLD_HOST=http://gpu-server:7680 mold run "a cat"

See the full CLI reference, configuration guide, and model catalog in the documentation.

Models

Supports 8 model families with 60+ variants:

Bare names auto-resolve: mold run flux-schnell "a cat" picks the best available variant.

See the full model catalog for sizes, VRAM requirements, and recommended settings.

Features

• txt2img, img2img, inpainting — full generation pipeline
• LoRA adapters — FLUX BF16 and GGUF quantized
• ControlNet — canny, depth, openpose (SD1.5)
• Prompt expansion — local LLM (Qwen3-1.7B) enriches short prompts
• Negative prompts — CFG-based models (SD1.5, SDXL, SD3, Wuerstchen)
• Pipe-friendly — echo "a cat" | mold run | viu -
• PNG metadata — embedded prompt, seed, model info
• Terminal preview — Kitty, Sixel, iTerm2, halfblock
• Smart VRAM — quantized encoders, block offloading, drop-and-reload
• Shell completions — bash, zsh, fish, elvish, powershell
• REST API — mold serve with SSE streaming, auth, rate limiting
• Discord bot — slash commands with role permissions and quotas
• Interactive TUI — generate, gallery, models, settings

Deployment

How it works

Single Rust binary built on candle — pure Rust ML, no Python, no libtorch.

mold run "a cat"
  │
  ├─ Server running? → send request over HTTP
  │
  └─ No server? → load model locally on GPU
       ├─ Encode prompt (T5/CLIP text encoders)
       ├─ Denoise latent (transformer/UNet)
       ├─ Decode pixels (VAE)
       └─ Save PNG

Requirements

• NVIDIA GPU with CUDA or Apple Silicon with Metal
• Models auto-download on first use (~2-30GB depending on model)