# mold
[](https://github.com/utensils/mold/actions/workflows/ci.yml)
[](https://codecov.io/gh/utensils/mold)
[](https://flakehub.com/flake/utensils/mold)
[](https://www.rust-lang.org)
[](https://nixos.wiki/wiki/Flakes)
Generate images and short video clips on your own GPU. No cloud, no Python, no fuss.
```bash
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
```bash
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.
<details>
<summary>Other install methods</summary>
### Nix
```bash
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
```bash
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](https://github.com/utensils/mold/releases).
</details>
## Usage
```bash
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 qwen-image-edit-2511:q4 "make the chair red leather" --image chair.png --image swatch.png
mold run ltx-video-0.9.6-distilled:bf16 "a fox in the snow" --frames 25
mold run "a cat" --expand # LLM prompt expansion
mold run qwen-image:q2 "a poster" --qwen2-variant q6 # Qwen-Image quantized text encoder
mold run flux-dev:bf16 "portrait" --lora style.safetensors # LoRA adapter
```
### Inline preview
Display generated images directly in the terminal (requires `preview` feature):
```bash
mold run "a cat" --preview
```
<p align="center">
<img src="docs/terminal-preview-example.png" alt="Generating the mold logo with --preview in Ghostty" width="720" />
<br/>
<em>Generating the mold logo with <code>--preview</code> in Ghostty</em>
</p>
### Piping
```bash
```
### Terminal UI (beta)
```bash
mold tui
```
<p align="center">
<img src="website/public/gallery/tui-generate.png" alt="mold TUI — Generate view with image preview" width="720" />
<br/>
<em>The TUI Generate view with Kitty graphics protocol image preview in Ghostty</em>
</p>
### Model management
```bash
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
```bash
# On your GPU server
mold serve
# From your laptop
MOLD_HOST=http://gpu-server:7680 mold run "a cat"
```
### Cloud GPU via `mold runpod`
Generate on a cloud GPU without managing pods yourself:
```bash
mold config set runpod.api_key <key> # one-time setup
mold runpod run "a cat on a skateboard" # creates pod → generates → saves to ./mold-outputs/
```
`mold runpod run` picks the cheapest available GPU, falls back across
datacenters if scheduling stalls, streams SSE progress over RunPod's
Cloudflare proxy, and leaves the pod warm for reuse on the next call.
See the [RunPod CLI guide](https://utensils.github.io/mold/deployment/runpod-cli)
for full subcommand reference (`doctor`, `gpus`, `list`, `create`, `stop`,
`delete`, `logs`, `usage`, …).
See the full [CLI reference](https://utensils.github.io/mold/guide/cli-reference), [configuration guide](https://utensils.github.io/mold/guide/configuration), and [model catalog](https://utensils.github.io/mold/models/) in the documentation.
## Models
Supports 11 model families with 80+ variants:
| **FLUX.1** | schnell, dev, + fine-tunes | Best quality, 4-25 steps, LoRA support |
| **Flux.2 Klein** | 4B and 9B | Fast 4-step, low VRAM, default model |
| **SDXL** | base, turbo, + fine-tunes | Fast, flexible, negative prompts |
| **SD 1.5** | base + fine-tunes | Lightweight, ControlNet support |
| **SD 3.5** | large, medium, turbo | Triple encoder, high quality |
| **Z-Image** | turbo | Fast 9-step, Qwen3 encoder |
| **Qwen-Image** | base + 2512 | High resolution, CFG guidance, GGUF quant support |
| **Qwen-Image-Edit** | 2511 | Multimodal image editing, repeatable `--image`, negative prompts |
| **Wuerstchen** | v2 | 42x latent compression |
| **LTX-2 / LTX-2.3** | 19B, 22B | Joint audio-video generation, MP4-first workflows |
| **LTX Video** | 0.9.6, 0.9.8 | Text-to-video with APNG/GIF/WebP/MP4 output |
Bare names auto-resolve: `mold run flux-schnell "a cat"` picks the best available variant.
See the full [model catalog](https://utensils.github.io/mold/models/) for sizes, VRAM requirements, and recommended settings.
### LTX Video
Current supported LTX checkpoints are:
- `ltx-video-0.9.6:bf16`
- `ltx-video-0.9.6-distilled:bf16`
- `ltx-video-0.9.8-2b-distilled:bf16`
- `ltx-video-0.9.8-13b-dev:bf16`
- `ltx-video-0.9.8-13b-distilled:bf16`
Recommended default today: `ltx-video-0.9.6-distilled:bf16`.
The `0.9.8` models pull the required spatial-upscaler asset automatically and
now run the full multiscale refinement path. mold keeps the shared T5 assets
under `shared/flux/...`, stores the `0.9.8` spatial upscaler under
`shared/LTX-Video/...`, and intentionally continues using the compatible
`LTX-Video-0.9.5` VAE source until the newer VAE layout is ported.
### LTX-2 / LTX-2.3
Current supported LTX-2 checkpoints are:
- `ltx-2-19b-dev:fp8`
- `ltx-2-19b-distilled:fp8`
- `ltx-2.3-22b-dev:fp8`
- `ltx-2.3-22b-distilled:fp8`
Recommended default today: `ltx-2-19b-distilled:fp8`.
This family is separate from `ltx-video`: it defaults to MP4, supports
synchronized audio, audio-to-video, keyframe interpolation, retake workflows,
stacked LoRAs, and camera-control LoRAs. The implementation is native Rust in
`mold-inference` with no Python bridge or upstream checkout requirement. CUDA
is the supported backend for real local generation, CPU is a correctness-only
fallback, and Metal is explicitly unsupported for this family. On 24 GB Ada
GPUs such as the RTX 4090, mold uses native staged loading, layer streaming,
and the compatible `fp8-cast` path for local FP8 runs rather than Hopper-only
`fp8-scaled-mm`. The native CUDA acceptance matrix is now validated across 19B
and 22B text+audio-video, image-to-video, audio-to-video, keyframe, retake,
public IC-LoRA, spatial upscaling (`x1.5` / `x2` where published), and
temporal upscaling (`x2`). The shared Gemma text assets are gated on Hugging
Face, so `mold pull` requires approved access to
`google/gemma-3-12b-it-qat-q4_0-unquantized`.
When you send source media through `mold serve`, the built-in request body
limit is `64 MiB`, which is enough for common retake and audio-to-video
requests without changing server config.
## Features
- **txt2img, img2img, multimodal edit, inpainting** — full generation pipeline
- **Image upscaling** — Real-ESRGAN super-resolution (2x/4x) via `mold upscale`, server API, or TUI
- **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
- **Qwen family encoder control** — selectable Qwen2.5-VL variants for Qwen-Image and Qwen-Image-Edit, with quantized auto-fallback when BF16 would be too heavy
- **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
| **NixOS module** | [Deployment: NixOS](https://utensils.github.io/mold/deployment/nixos) |
| **Docker / RunPod** | [Deployment: Docker](https://utensils.github.io/mold/deployment/docker) |
| **mold runpod CLI** | [Deployment: RunPod CLI](https://utensils.github.io/mold/deployment/runpod-cli) |
| **Systemd** | [Deployment: Overview](https://utensils.github.io/mold/deployment/) |
## How it works
Single Rust binary built on [candle](https://github.com/huggingface/candle) for the in-tree model families. LTX-2 now runs through the native Rust stack in `mold-inference`, so the full model surface stays in Rust with no libtorch dependency.
```
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)