mold

Appearance

Getting Started ​

mold began as a single-binary command-line tool for generating AI images and video on your own GPU, and it remains CLI-native. No cloud API, Python environment, or heavyweight runtime is required.

The command line is the stable product foundation: core workflows accept predictable arguments, stdin, files, and environment variables; return useful stdout, stderr, exit status, and machine-readable output, and work the same for a person, shell script, CI job, or agent. Desktop, web, TUI, iPhone, REST/SSE, and MCP clients extend that engine rather than hiding it.

How It Works ​

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

mold tries to connect to a running mold serve instance first. If no server is available, it falls back to local GPU inference; auto-downloading the model if needed.

Tea house, generated with FLUX Dev Q4

Built for automation ​

Mold composes with ordinary Unix tools:

bash
# Prompt on stdin, image bytes on stdout
echo "a cinematic city at night" | mold run flux-schnell | viu -

# Reproducible batch work from a script or CI job
mold run flux-dev:q4 "product photo" --seed 42 --output artifact.png

# Point the same CLI or an agent at a remote GPU
MOLD_HOST=http://gpu-server:7680 mold run "a storyboard frame"

For tool-calling agents, use mold mcp, the OpenClaw skill, or the documented REST/SSE API.

Quick Start ​

bash
# Install (auto-detects your GPU)
curl -fsSL https://raw.githubusercontent.com/utensils/mold/main/install.sh | sh

# Generate your first image
mold run "a sunset over mountains"

# That's it: mold shows the complete download before the first pull

For the latest rolling CLI build from main, add MOLD_CHANNEL=nightly on the sh side of the install pipe.

What You Get ​

  • Broad model support: FLUX.1, SDXL, SD 1.5, SD 3.5, Z-Image, Flux.2 Klein and Dev, Qwen-Image, Qwen-Image-Edit, Wuerstchen v2, LTX Video (0.9.x, 2, 2.3, and 2.5), Wan 2.1/2.2, and MiniMax H3
  • txt2img, img2img, multimodal edit, inpainting, ControlNet: all in one binary
  • Image upscaling: Real-ESRGAN super-resolution (2x/4x) via CLI, server API, or TUI
  • Pipe-friendly: mold run "a cat" | viu - just works
  • Client-server: run the GPU part on one machine, generate from anywhere
  • Native apps: a local/multi-host desktop studio plus remote-only iPhone and Android companions
  • Prompt expansion: short prompts become detailed via local LLM
  • LoRA adapters: apply fine-tuned styles across FLUX, Flux.2, LTX-2, SD1.5, SD3, SDXL, Qwen-Image, Qwen-Image-Edit, Wan, and Z-Image
  • PNG metadata: generation parameters embedded for reproducibility

Requirements ​

  • For practical local generation: an NVIDIA GPU with CUDA or Apple Silicon with Metal. CPU correctness paths and remote-client-only use do not require a local GPU.
  • Models auto-download on first use (~2 GB to 50+ GB depending on model and optional assets)

Next Steps ​