Mold 发布本地 AI 图像、视频与 3D 生成工具链

Local AI image and video generation on your own GPU. Mold supports NVIDIA CUDA and Apple Silicon Metal, with a CLI, native desktop app, web studio, mobile companions, Discord bot, and REST/SSE API built on the same engine.

Documentation · Models · Desktop guide · API

Install

Stable release:

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

Nightly CLI from the latest published main build:

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

The installer selects a compatible build and verifies its checksum. Linux x86_64 clients without a visible NVIDIA GPU receive a GPU-free CLI. To use only remote GPU hosts, including from a machine with NVIDIA hardware:

curl -fsSL https://raw.githubusercontent.com/utensils/mold/main/install.sh | MOLD_BACKEND=cpu sh
MOLD_HOST=http://gpu-server:7680 mold run "a cat"

The CPU archive needs no NVIDIA driver or CUDA libraries. Local GPU generation requires a CUDA build (or Metal on macOS). See the installation guide for Nix, Arch, Windows, Android, and source builds. GH200, GB200, and GB300 require future linux/arm64 artifacts and are unsupported.

Mold no longer publishes new versions to crates.io. Existing registry versions are historical; use GitHub releases, Nix, Docker, AUR, or a source build for current versions.

Quick start

# Generate with the default model
mold run "a cat riding a motorcycle through neon-lit streets"

# Choose a model and reproducible seed
mold run flux-dev:q4 "a sunset over mountains" --seed 42

# Edit an image
mold run qwen-image-edit-2511:q4 "make the chair red" --image chair.png

# Edit from reference images (repeat --reference; order is semantic)
mold run flux2-klein:bf16 "the woman from image 1 wearing the glasses from image 2" \
  --reference person.jpg --reference glasses.jpg

# Qwen Image 2.1: up to 10 ordered references, a transparent cut-out, or 6-step turbo
mold run qwen-image-2.1 "put the jacket from image 1 on the person in image 2" \
  --image jacket.png --image person.jpg
mold run qwen-image-2.1 "a red paper lantern with a gold tassel" --transparent -o lantern.png
mold run qwen-image-2.1-turbo "a lighthouse on a basalt cliff at dusk, oil painting"

# Generate video
mold run ltx-video-0.9.6-distilled:bf16 "a fox in the snow" --frames 25

# Turn a photo into a 3D mesh
mold run hunyuan3d-mini-turbo --image chair.png -o chair.glb
mold run hunyuan3d-2mv-turbo --front chair-front.png --left chair-left.png --back chair-back.png -o chair.glb

# Upscale a Library video as a durable framewise job
mold video-upscale create clip.mp4 --wait

# Launch the web studio and API
mold serve

For API-only servers, set web_ui_enabled = false in config.toml (or MOLD_WEB_UI_ENABLED=false) and restart. The interface defaults on; disabling it leaves every API route available. NixOS: services.mold.webUi.enable = false;.

Models download automatically on first use. Generated media is saved locally with prompt, model, seed, and generation metadata. Framewise video upscale also needs the host codec bridge: Nix packages and CUDA containers include it, while raw binary installs must provide ffmpeg and ffprobe on PATH before the server advertises that feature.

What it supports

• Models: FLUX.1, Flux.2, Stable Diffusion, Z-Image, Qwen-Image, Qwen Image 2.1, Wuerstchen, LTX Video, Wan, MiniMax H3, and Hunyuan3D. See the model catalog for variants and hardware requirements. Wan's 1.3B BF16 and 5B Q8/FP16 paths are performance-qualified on Apple Metal as well as CUDA; fp8-scaled Wan checkpoints remain CUDA-only. MiniMax H3's compact runtime is supported on SM89 CUDA and Apple Metal; Metal admission checks each request against live unified-memory headroom. FL2VA requires at least one boundary frame: first, last, or both, with roughly 1,000 prompt tokens with paired frames and charging their additional memory. Queued FL2VA jobs preserve first/last endpoint roles through durable storage and replay, including all supported Turbo tiers. Forced-local H3 execution accepts one FL2VA request; batches, sequences, and Ref2VA reference uploads require the server route. Qwen Image 2.1 (non-commercial Qwen Research License) edits from up to ten ordered references, renders transparent PNG/WebP cut-outs, runs native 2K presets, and ships BF16, INT8, FP8, GGUF and 6-step turbo tiers; its CUDA fast path is 2.5x faster at 1024² and 6x at 2K than v0.32.
• Images: text-to-image, image editing, inpainting, ControlNet, LoRA, identity photos, reference-image prompting, transparent backgrounds, prompt expansion, and upscaling, as PNG, JPEG or still WebP.
• Video and audio: text/image-to-video, sequences, clip continuation, lip dub, text-to-audio, and MP4 output with generated audio.
• 3D: single-image and named multiview-to-mesh with Hunyuan3D 2.0 and 2.1, automatic background removal, optional highlight and lighting removal, and Hunyuan3D Paint PBR materials in CUDA builds. Shape transformers can be locally derived as qualified FP8/GGUF tiers with mold quantize, which registers the derived tier on the host that made it. Results are published to the Library as binary glTF with a rendered poster tile, exportable as OBJ, an OBJ+PBR ZIP, STL, or PLY, or shared as a turntable GIF, APNG, or WebP. Durable text-to-3D, supplied-mesh texturing, and mesh-rebuild workflows remain available through the API/CLI. Apps generate 3-D objects directly from supplied images in the main generation screen. Painted prints expose their base-color, metallic-roughness, and normal maps as independent, digest-checked Library downloads on web, desktop, and mobile.
• Multiple machines: connect local, LAN, Tailscale, and RunPod hosts, then route work and browse one combined Library.
• Organization: title, favorite, tag, collect, restore, and manage prints across the desktop and web apps.

Curated models use the same short manifest title throughout the apps and mold list, such as FLUX.1 Dev Q4. The runnable ID (flux-dev:q4) remains available for commands, saved settings, and API requests. Model files keep their original names; third-party models keep their provider titles.

Model weights keep their own licenses. See each model page for terms and current platform support.

Mold Studio

On a Mac, use Mold Studio. It is the native macOS app: signed, notarized, self-updating through Sparkle, with mold's own Metal engine built in, and it talks to any mold serve on your network too. Generate, a merged multi-machine Library, Queue, Models and Machines, all in one native window. Queue jobs open into grouped details and live previews; Discover starts with manifest models. Generate confirms submissions, offers explicit random or fixed seeds, and lets you resize the prompt editor and settings inspector. Native playback settings control autoplay/repeat, and oversized GUI input pictures are fitted automatically while library originals remain unchanged. The Generate controls stay at the bottom; the Library can create a collection from selected prints and hide collection members from general browsing across all their machine copies. The native iOS companion shares collection hiding. Mold Studio for Mac guide

Download Mold Studio for Mac (Apple Silicon, macOS 26+)

It installs as Mold Studio.app beside the older Tauri app (Mold.app), so both can live in /Applications while you move over. The Tauri app is now the legacy Mac download — keep it only for a Mac that cannot run macOS 26: Download Mold Desktop for macOS (legacy).

On Windows and Linux the Tauri desktop app remains the default. It puts New image, Queue, My images, Styles, Machines, and Settings in one window, in plain words, with six themes, for local and remote generation, and pairs with the iPhone and Android companions. Explore the desktop app

Download Mold for Windows (x86_64) — a self-signed NSIS installer. The published build is CPU / remote-hosts only; see the desktop guide for the CUDA recipe. Verify and explicitly trust the release's mold-windows-self-signing.cert.cer before installing; the certificate is not publicly trusted and does not suppress SmartScreen on its own.

Linux desktop builds are source/CI distributions for now — nix build .#mold-desktop or the devshell's desktop-build CUDA AppImage. See the desktop guide.

On iPhone and iPad, Mold Studio Companion is the native app beside Mold Studio for Mac: pair by scanning the Mac's code, generate on your machines, and follow renders from the Lock Screen. Generation options adapt to large text, and the full-screen viewer keeps its media actions accessible. It is in TestFlight while it is new. Mold Studio for iPhone & iPad guide

The native iOS companion uses concise render notifications with its app icon; tapping opens the finished print from the background or a cold launch, with notification activation completed on the main thread. Library long-press and drag previews retain their thumbnail loader, and source selection prefers a reachable machine copy; the notification icon keeps the same authored colors in light and dark appearances. Its Live Activity uses the system background material with matching semantic text, a full-width progress row, and a separate machine/queue footer. Generate Options marks the selected aspect with a checkmark; attaching the first/start frame chooses the closest supported aspect, and a closing frame preserves it. Options shows proportionate aspect icons, source-image fitting (centered Crop to fill by default), and an explicit Random seed default.

The native iOS companion offers Open Settings recovery for denied Photos, camera, nearby-network, notification and Live Activity access. Photos saving uses add-only authorization and safely writes pictures and videos; photo imports use the system picker. See the native permission audit.

The native iOS Library keeps its offline notice compact and offers scrollable details with every full machine name, preserving room for saved prints at large text sizes. Decorative Library badges stay within their tiles; the full machine names remain in accessibility labels and Info.

The native iOS Models screen offers visible per-model Unload and Unload All Models controls for the selected server; unloading keeps downloaded files.

Native iOS development: nix develop -c companion-dev watches Swift sources, builds and relaunches in Simulator. companion-run launches once; see the native iOS developer guide for simulator and build-directory overrides.

Android uses the same remote-only Mold Studio mobile interface. Download the signed universal nightly APK directly; there is no zip to unpack:

Download nightly Android APK · Android installation guide

More ways to create

Preview generations directly in supported terminals:

mold run "a cat" --preview

Inline image generation in Ghostty with --preview

Run the engine where the GPU lives and connect from another machine:

mold serve                                      # GPU machine
MOLD_HOST=http://gpu-server:7680 mold run "a cat"  # laptop

For access outside your network without opening the GPU host's inbound ports, use the optional HTTPS reverse relay. Run mold relay connect --transport aws beside an authenticated server; every client uses the resulting HTTPS machine address with its existing API key or pairing flow. The trusted Lambda gateway can see traffic and publishes one machine per address. Authenticated servers allow their browser shell/assets to load before sign-in; API and media access still require their existing permissions.

--offload also applies to remote renders and durable sequences on GPU hosts. Add --no-save to keep one render out of a server's Library; the host still publishes the print and moves it straight to trash, so mold trash restore gets it back until retention sweeps it. It applies to renders a server performs — a local render has no Library, and refuses the flag rather than ignoring it.

Pass --fit beside --image when the source and the canvas disagree. Without it the picture decides the canvas; with it the canvas is what you asked for and the picture is resampled onto it — crop-fill trims the edges, pad-fit adds black borders, lanczos-resize stretches.

mold run flux-dev:q4 "a lighthouse at dusk" --image wide.png --fit crop-fill --width 1024 --height 1024

A sequence of several clips is scripted, not composed in an app:

mold chain validate shot.toml
mold run --script shot.toml --output walk.mp4
mold jobs list

A 3-D render can be one shot or a durable workflow. A workflow keeps every stage — the picture it starts from, its matted and delighted copies, the shape, the paint — as its own retained artifact, reports each stage as it changes state, and resumes after a restart. It lives on the machine that runs it:

mold mesh-workflow create --prompt "a small ceramic fox" --texture --follow
mold mesh-workflow create --mesh chair.glb --image chair-albedo.png
mold mesh-workflow list

Find weights to install, and watch them arrive:

mold search "anime style" --kind lora
mold pull flux-dev:q4
mold downloads watch

See the remote workflow and RunPod guides. Use mold queue to manage remote work and mold library to browse and organize the host's prints, including mold library source-media to recover the conditioning image a print was made from and mold trash delete to remove one permanently. Retained sources survive queue cleanup and remain available while any referencing library output exists, including in the trash. Matching retained input sets from newly completed jobs share storage while keeping per-job provenance. Native Mac and desktop library copies also retain those sources on supported destinations. Native Mac mirrors accept missing archive-only job IDs and generation durations, a short version matching the same version with a build suffix, and an alpha-channel fact derived from identical output bytes when absent on the source. The same compatibility check compares the gallery listing with its stable source archive, so completion bookkeeping does not falsely report a changed print. When an older archive omits scheduler or transparency settings, the source output is downloaded and its digest verified before its embedded recipe can corroborate those fields. Output bytes and all generation settings must still match; conflicting recorded provenance is refused. Retrying Sync All repairs retained inputs on compatible existing copies without duplicating their outputs.

Native Mac Save Locally and Sync All count a verified finished print as saved when its source explicitly reports legacy inputs as unavailable. The sync summary reports unavailable original inputs separately; it does not claim they were retained. Existing copies still require matching output bytes and recipes. Corrupt or unauthenticated input archives, changed outputs, and unsupported source transfer contracts remain failures.

Native Mac shipping builds include the reviewed Metal H3 runtime. Remote generation follows the selected host’s capabilities. Reused native Mac references show bounded previews and survive ordinary control edits and repeated submissions; relaunch verifies their byte-free origin locator before enabling Generate.

Native Mac local authoring inputs are saved privately alongside the prompt and scalar settings, including first/last frames and parked attachments. Unresolved retained references still require their original machine to validate them. Save failures are visible; missing or corrupt saved inputs block Generate until you reattach what you need and explicitly choose Use current inputs.

Before importing a source-bearing copy, clients check destination readiness. Windows local destinations currently cannot receive retained inputs, so those copies are refused before creating a local library output. Source-free copies remain supported, and Windows clients can recall retained sources from a supported remote machine. To install Mold's Agent Skill for supported coding agents, run:

mold skill install --detected

The installed bundle uses each agent's native metadata and discovery contract, with a concise router, safety guidance, tested examples, and the prompting corpus: a shared guide, one complete base guide per manifest family (prompt style, syntax, generation context, examples, pitfalls, and that family's CLI examples), task leaves for the distinct H3, Wan, and LTX-2 grammars, and model leaves for checkpoints with quirks of their own. The corpus in crates/mold-core/src/prompting/ is also what mold expand, mold remix, --expand, the app Expand and Remix actions, and the MCP expand_prompt / remix_prompt tools hand to the LLM, together with the exact model, canvas, frame count, fps, and ordered references, so agents and the expander follow one set of rules. Hunyuan3D's base guide is the one that tells an agent NOT to write a prompt.

macOS Metal memory

Use mold system metal-memory status to inspect this Mac, or mold gpu list --json for a running host. Explicit root-only set <MiB> / reset commands support an optional boot policy with --persist. See the Metal memory guide for budget accounting, local-only administration and rollback semantics. Z-Image's Metal whole-decode path retries with tiles on memory errors and preserves the eager CPU fallback. CPU/CUDA decode ordering is unchanged; see the VAE qualification record.

Project

Mold is a Rust workspace built on candle. The documentation covers the CLI, configuration, deployment, and HTTP API.

Core contributors: James Brink and Jeffrey Dilley.

Licensed under the MIT License. Third-party components and model licenses are listed in THIRD_PARTY_NOTICES.md and the model documentation. InsightFace identity weights require separate acceptance and are limited to non-commercial research use.

Model checksums are verified when files are downloaded. Complete installed models queue and switch without full checksum scans, including after restart. To check existing bytes explicitly, run mold info MODEL --verify. Normal loading still checks file sizes and formats; it does not guarantee detection of same-size corruption.

The public website privacy policy describes Google Analytics on the documentation website. Analytics loads automatically without a popup; this integration is not included in Mold apps or servers.

Native iOS and macOS Libraries offer All Media, Photos, Videos and 3D filters within each shelf. iOS Select supports finger sweeps across a range, including edge scrolling, and keeps the viewport stable when selecting.

Native macOS Settings ▸ Remote Access, beside Machines, offers Pair your phone in the existing first section, alongside the connection address and paired-device controls. Scan its QR in the native iOS app or Tauri mobile app. Saved machines advertise only routes their listener actually serves: phone clients prefer LAN, then Tailscale, with HTTPS proxy fallback after an instance-bound credential-free proof. This Mac's loopback listener advertises only its prepared managed HTTPS origin. An existing paired machine keeps its identity and credential as routes change. For an independently hosted server, configure MOLD_PUBLIC_URL to advertise its actual public HTTPS relay origin.

Automatic roaming requires a server-minted paired credential. Operator API keys remain tied to the explicitly saved address; pair once to enable route learning. Probes expose a stable digest tag, so arbitrary operator keys never participate in anonymous route proofs. Plain HTTP still requires a trusted network.

Native iOS queue cards use concise curated model names and open full job details on tap, with state-aware per-job controls. Generate exposes searchable per-machine Prompt History with prompt-only recall.

Native iOS and macOS generation controls support ordered MiniMax H3 image/video/audio references, Hunyuan3D named views, and Wan/MiniMax boundary frames. See the native reference parity audit and native app guides for limits and media formats.

Media export controls

Native iOS and Tauri offer clip GIF exports with Loop/Bounce, Forever/Once, size, frame rate and, on supporting hosts, an extra boundary pause. Set the pause to 0 ms to add no hold to a continuous source loop. GIF frame timing remains FPS-derived; a pause does not repair a discontinuity in the source. Native iOS also exports host-advertised mesh formats, turntables and texture sidecars to Share, Files or its Mold folder, with GIF delivery to Photos. See native export coverage.

Image inputs above a model’s processing size are prepared automatically. Native reference strips accept ordinary oversized pictures; processing pixel budgets are separate from bounded image ingestion and request transport limits. Large identity photographs are proportionally reduced to the face detector’s ingestion envelope, and original files remain untouched.

Queue rows show the owning machine’s sealed conditioning images across models. Job Details shows every ordered reference separately with its role, including identity photos, named views, masks, control images and boundary frames. Audio/video references are listed by kind; unavailable previews are disclosed. Inputs remain separate from live denoise previews and are fetched through authenticated routes, including work submitted from another device. Older servers retain their singular source preview.

Error messages

Errors use concise summaries instead of internal traces. Memory refusals show estimated requirements, the available budget and the shortfall in decimal GB/MB; a shared-memory Mac identifies the shared pool. These are estimates on the generation machine, not a promise that freeing exactly that amount guarantees a render. Full server diagnostics remain in logs and durable records. Both native apps share MoldClient error presentation, including older-server replies; local failures are recorded in OSLog with privacy-protected details.

Queue model downloads

A missing-model held job offers Download and Retry in Queue or Job Details. The job shows Starting, download-queue status, live bytes/progress, license review, reconnection and failures in place. Closing details does not stop recovery. The download runs on the job’s owning machine; retry occurs only after every returned download ticket succeeds and the original held job and server identity are revalidated. Failed or cancelled downloads leave the job held. Global queue pause stays in effect. Cancelling a job does not cancel a model download that other jobs may need.

Reuse input media

Reuse settings restores retained opening and closing frames into editable wells, preserving keyframe indices and ordered image references. Audio, source/continuation video, identity photos and control images restore with their saved settings where the selected recipe supports them. Continuation overlap and authored reference strength (including zero) are preserved. Older prints without recorded reference strength keep the server default. The original machine must retain the inputs; missing, damaged or oversized inputs are disclosed before generation. A new attachment or explicit removal wins over a late download, and restored wells never have a hidden archive fallback that can revive removed media.