mold

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

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

Nix (recommended)

# Run directly — no install needed
nix run github:utensils/mold -- run "a cat"

# Or add to your system
nix profile install github:utensils/mold

From source

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

Usage

# Generate an image
mold run "a sunset over mountains"

# Pick a model
mold run flux-dev:q4 "a turtle in the desert"
mold run sdxl-turbo "espresso in a tiny cup"
mold run dreamshaper-v8 "fantasy castle on a cliff"

# Reproducible results (the logo above was generated this way)
mold run z-image-turbo:bf16 "A minimal modern logo for 'mold' on a solid black background. A stylized casting mold shape formed from smooth gradient lines transitioning from cyan to magenta. The negative space inside the mold reveals a glowing latent grid pattern suggesting AI diffusion. Bold lowercase 'mold' typography below in clean sans-serif. Flat vector style, no photorealism" --seed 1337

# Custom size and steps
mold run "a portrait" --width 768 --height 1024 --steps 30

Piping

Mold is pipe-friendly in both directions. When stdout is not a terminal, raw image bytes go to stdout and status/progress goes to stderr.

# Pipe output to an image viewer
mold run "neon cityscape" | viu -

# Pipe prompt from stdin
echo "a cat riding a motorcycle" | mold run flux-schnell

# Chain with other tools
cat prompt.txt | mold run z-image-turbo --seed 42 | convert - -resize 512x512 thumbnail.png

# Pipe in and out
echo "cyberpunk samurai" | mold run flux-dev:q4 | viu -

Output metadata

PNG output embeds generation metadata by default, including prompt, model, seed, size, steps, and a JSON mold:parameters chunk for downstream tools.

# Disable metadata for one run
mold run "a cat" --no-metadata

# Disable metadata globally via environment
MOLD_EMBED_METADATA=0 mold run "a cat"

In ~/.mold/config.toml:

embed_metadata = false

Image-to-image

Transform existing images with a text prompt:

# Stylize a photo
mold run "oil painting style" --image photo.png

# Control how much changes (0.0 = no change, 1.0 = full denoise)
mold run "watercolor" --image photo.png --strength 0.5

# Pipe an image through
cat photo.png | mold run "sketch style" --image - | viu -

Inpainting

Selectively edit parts of an image with a mask (white = repaint, black = keep):

mold run "a red sports car" --image photo.png --mask mask.png

ControlNet (SD1.5)

Guide generation with a control image (edge map, depth map, etc.):

mold pull controlnet-canny-sd15
mold run sd15:fp16 "a futuristic city" --control edges.png --control-model controlnet-canny-sd15

Scheduler selection

Choose the noise scheduler for SD1.5/SDXL models:

mold run sd15:fp16 "a cat" --scheduler uni-pc        # Fast convergence
mold run sd15:fp16 "a cat" --scheduler euler-ancestral # Stochastic

Batch generation

Generate multiple images with incrementing seeds:

mold run "a sunset" --batch 4    # Generates 4 images: seed, seed+1, seed+2, seed+3

Manage models

mold pull flux-schnell:q8    # Download a model
mold list                    # See what you have
mold info flux-dev:q4        # Model details + disk usage
mold rm dreamshaper-v8       # Remove a model

Remote rendering

Run mold on a beefy GPU server, generate from anywhere:

# On your GPU server
mold serve

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

Models

FLUX (best quality)

SDXL (fast + flexible)

SD 1.5 (lightweight)

SD 3.5

Z-Image

Wuerstchen v2 / Flux.2 / Qwen-Image (alpha, improving on CUDA/MPS)

Warning: These model families are still in active alpha development. Results vary by backend and may be better on CUDA than Apple Silicon (MPS/Metal). Use FLUX, SDXL, SD1.5, SD3.5, or Z-Image for production use.

Bare names resolve by trying :q8 → :fp16 → :bf16 → :fp8 in order. So mold run flux-schnell "a cat" just works.

Server API

When running mold serve, you get a REST API:

# Generate an image
curl -X POST http://localhost:7680/api/generate \
  -H "Content-Type: application/json" \
  -d '{"prompt": "a glowing robot"}' \
  -o robot.png

# Check status
curl http://localhost:7680/api/status

# List models
curl http://localhost:7680/api/models

# Interactive docs
open http://localhost:7680/api/docs

Shell completions

source <(mold completions bash)    # bash
source <(mold completions zsh)     # zsh
mold completions fish | source     # fish

Requirements

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

AI Agent Skill

Mold ships with an AI agent skill that teaches AI assistants how to use the CLI for image generation. This lets agents generate images on your behalf using natural language.

Claude Code

The skill is automatically available when working in the mold repo. To use it in other projects, copy the skill directory:

# Copy to your project (project-scoped)
cp -r path/to/mold/.claude/skills/mold .claude/skills/

# Or install globally (available in all projects)
cp -r path/to/mold/.claude/skills/mold ~/.claude/skills/

Then use it via /mold a cat on a skateboard or let Claude invoke it automatically when you ask to generate images.

OpenClaw

Copy the skill to your OpenClaw workspace:

cp -r path/to/mold/.claude/skills/mold ~/.openclaw/workspace/skills/

Or install directly from the repo:

git clone --depth 1 --filter=blob:none --sparse https://github.com/utensils/mold.git /tmp/mold-skill
cd /tmp/mold-skill && git sparse-checkout set .claude/skills/mold
cp -r .claude/skills/mold ~/.openclaw/workspace/skills/
rm -rf /tmp/mold-skill

The skill format is compatible with both Claude Code and OpenClaw (both use SKILL.md with YAML frontmatter).

How it works

Mold is a single Rust binary built on candle — a pure Rust ML framework. No Python runtime, no libtorch, no ONNX. Just your GPU doing math.

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