iPhone App

Mold for iPhone is a first-class remote studio for a Mold server. The phone does not run AI models itself: generation, model storage, downloads, queueing, and gallery files stay on one or more remote Mold hosts while the app provides an iPhone-native control surface.

Mold requires iOS 17 or later. The current build is distributed to invited internal and external TestFlight groups; there is not yet a public App Store listing.

Connect a host

Start a reachable server on the GPU machine. Use an API key whenever the server is reachable beyond localhost:

MOLD_API_KEY='choose-a-long-secret' mold serve --bind 0.0.0.0 --port 7680

Open Machines on the iPhone and use one of these paths:

• Recommended: open Settings → Mobile pairing in Mold desktop or web, confirm the LAN, MagicDNS, or HTTPS address the phone can reach, tap Create pairing code, then tap Scan pairing code on iPhone. The code is single-use, expires after two minutes, and contains no API key; the redeemed key goes directly into the iOS Keychain.
• Tap Discover nearby to browse _mold._tcp services on the current LAN. Allow Local Network access when iOS asks.
• Enter an IP address such as 192.168.1.10. Mold adds http:// and port 7680 when they are omitted.
• Enter a DNS hostname, HTTPS URL, or Tailscale MagicDNS name.

For manual entry, paste the same API key, then tap Test and save. Saved host metadata remains in the app's local storage, while the API key is kept separately in the iOS Keychain.

Tailscale

Tailscale is the simplest way to reach a private GPU host away from its LAN:

1. Join the iPhone and GPU host to the same tailnet.
2. Keep mold serve listening on an address the Tailscale interface can reach.
3. Add the host by its MagicDNS name, for example plato or plato.example-tailnet.ts.net. Mold supplies port 7680 for a bare name.
4. Include https:// and an explicit port only when your own reverse proxy provides TLS.

Bonjour discovery is LAN-local; Tailscale hosts are normally added by name. Mold uses the installed Tailscale network and does not manage tailnet login, ACLs, DNS, or certificates.

Create

Choose where the print develops and one of the installed generation models. With a single connected machine the Host control works as it always has. Once two or more connected machines are reachable it also offers:

• Auto: the least busy machine that already has the selected model.
• Most capable: the strongest GPU that has it: CUDA before Metal, then VRAM, then queue depth.

An optional Title above the prompt names the print: it is embedded in the saved metadata, folded into the output filename as a slug, and shown across every Library. Batch siblings and prepared variations inherit it, and Use as prompt restores it from a print.

Directly beneath it, File under files the print as you make it, as two rows:

• Tags: a dashed chip offers the title's own slug (tap it away and it stays away), and the sheet takes typed names and suggests the tags your machines already use, with counts.
• Collection: the collection whose name matches the title is pre-selected, never created; the sheet offers None, every collection merged across your machines, and an inline New collection… that only records the name until the print develops.

A line under the rows previews the filename the print will land as. Both rows disappear when the machine the print is routed to cannot organize its library, and Settings ▸ Library ▸ Tag new prints with their title turns the title chip off without touching prints you already made.

Under either policy the model list is the union of every reachable machine's installed models, and a model that is not on all of them is tagged with the machine that has it. Mold ranks the reachable machines from their cached model, queue, and GPU telemetry and freezes the winner before queueing, so recovery and prepared expansion both stay on the exact machine that ran the work. If no machine can run the print, nothing is queued and the message names each machine. Your choice is remembered; it falls back to the browsed machine while only one is reachable.

The form adapts to the selected model family and uses the same request contract and model defaults as desktop. The primary controls stay on the main screen; deeper options open in a full-screen More settings sheet, and prompt style presets compose at submit without rewriting your prompt text. A ↺ Reset beside the More settings trigger restores every generation setting to the selected model's defaults, keeping your prompt, model choice, and any prepared batch.

The mobile composer includes:

• a directly editable Batch count, Batch 1 quick prompt expansion with undo, compact/paged Batch N prepared-variation review, and remote prompt history;
• local templates that can be saved, searched, sorted, renamed, loaded, and deleted;
• batch generation as independent queued jobs, each with its own progress and Cancel action;
• source images and fit policies, Qwen edit target/reference images, masks, ControlNet, LoRA stacks and trigger words;
• scheduler, CFG++, steps, guidance, output format, and post-generation upscaling where the selected family supports them;
• validated video frames/FPS, audio, camera motion, source media, keyframes, retake, Continue a video (extending an existing clip on models that advertise it, including capable Wan checkpoints; resolution and fps stay locked to the source clip), LTX-2 pipeline/spatial/temporal controls, and optional STG, CFG-rescale, modality-scale, and guidance-skip overrides;
• a host-aware memory estimate before submission.

For video-only LTX-2 community checkpoints, Generate audio is disabled with an explanation when the connected server reports that the installed files lack an audio VAE or vocoder. Source-image video generation remains available. Empty guidance-override fields preserve the pipeline defaults. Invalid block lists or numeric ranges stay in the More settings sheet with inline feedback and cannot queue a request; Reset advanced, templates, and Library reuse share the same saved override state as desktop.

Identity photos (PuLID)

An identity photo conditions the print on a person's face: the render keeps that likeness while the prompt decides everything else. The photo itself is never composited into the output, and (unlike a source image) it is never cropped or fitted to the canvas. It is a reference, not a composition input.

An Identity well sits in the Create form beside the source wells whenever the selected model and the machine you are developing on both support it. When they do not, the control is not there at all rather than present and disabled. Tap it to pick a PNG or JPEG (at most 16 MiB, 8192 px per side, 32 MP) with the usual iOS photo/camera picker. Identity conditioning is offered only for the identity-qualified checkpoints on a server built with the feature, and it cannot be combined with a LoRA or an img2img source image; see Identity Photos (PuLID) for the full rule and the one-time InsightFace licence acceptance.

Switching to a model that cannot use an identity photo does not throw yours away and does not stop you developing: the photo is parked, the request goes out without it, and the well comes back with the photo still in it when you select a qualified model again.

Two knobs live in the More settings sheet, count toward its badge, and clear with Reset. Both stay absent from the request until you touch them, so the server's own defaults keep applying:

• Identity strength: how strongly the face is held, 0.0–3.0 (default 1.0). Higher preserves the likeness; lower lets the prompt reshape it.
• Identity start step: the first denoise step the face is applied at (default 0, always fewer than the print's step count). Delaying it lets the composition settle before the likeness is pinned.

If the combination cannot be submitted (a photo alongside a LoRA or a source image, a knob set with no photo, an oversized or unsupported file) the reason reads inline beside the control and Develop stays blocked. Prepared Batch N siblings inherit the photo and both knobs (and the reviewed card names the same reason on its own Develop), and changing the photo stales reviewed prompt work exactly as changing a source image does.

Under Auto or Most capable, an identity print is only ever sent to a machine that advertises identity support for that model itself; the model list is merged across your machines, so the one you staged the photo against is not necessarily the one that develops it. If the machine that was chosen cannot hold the face, or is running a Mold old enough to ignore the photo, Mold says so and queues nothing rather than returning a print of someone else.

Choose resolution through proportionally drawn shape tiles (the canonical families (1:1, 5:4, 4:3, 3:2, 16:9, 21:9 and their portrait twins) the selected model can actually express, plus Source when an image is attached) then a Size row that lists that family's authored sizes as exact pixels with their megapixels underneath. Explicit custom dimensions stay available. One status line names the canvas and why it holds that size (Matches source, Follows source, Model default, Manual), and the badge beside it always agrees with that sentence. Seed has separate Random and Fixed modes, including one-tap reuse of the last generated seed.

When several jobs are submitted, the Queue section keeps them visible while you continue composing. Mold limits simultaneous streams so generation does not starve gallery or download requests.

Queue rows on a machine's own screen swipe like any iOS list: drag right to left to reveal Cancel, or keep going and a full swipe cancels the job. Revealing the buttons is the first step and tapping one is the second, so nothing is cancelled by a single flick. Every action is also on the ⋯ button at the end of the row, so VoiceOver and a hardware keyboard reach it without the gesture. Machines that support reordering also offer To back, which sends the job to the end of the line. Tap the row itself to see the whole job; its prompt, its settings, where it is in line, its live preview while it runs, and the full reason if the machine has parked it.

For model families that stream live latent previews (FLUX.1, Flux.2, Z-Image, and Wan 2.1/2.2), the active print develops right on the Create screen: the preview sharpens as denoising progresses under a thinning film-grain wash, in a bed matching the print's aspect ratio. Hosts without previews keep the plain status line.

Batch is also the expansion count. Batch 1 freezes the selected host through the next Develop while keeping quick undo. Batch 2 or greater first requests exactly N distinct non-empty prompts from that host and opens an inline workspace where you can edit, intentionally remove, regenerate, refresh, or discard them before anything is queued. Above eight, the workspace shows a compact first-eight summary and bounded pages on demand. Queue acceptance immediately restores the composer so another batch can be prepared while the earlier one runs. A single reviewed set has a 10,000-variation memory-safety ceiling, but you can keep queueing additional sets. Reducing two prompts to one requires confirmation. Changes to the source prompt, model, family, Batch count, host endpoint, Keychain key, or server identity preserve the reviewed work with a specific stale reason and block Develop until you choose Refresh or Discard.

The same frozen host is used for expansion-model pulls, source preprocessing, and every sibling. Missing-model recovery stays inline with Connecting, Starting, Queued, percentage/bytes/files/ETA Pulling, Ready, and failure or cancellation retry states. The original inputs and route remain one immutable recovery record, and an attempt lease prevents a newer Models credential from redirecting the pull. Compatible Models work already in Starting is joined; every terminal, stale, superseded, or aborted attempt releases the lease before Retry reacquires the same frozen route. Editing/removing reviewed work cancels a pending replacement. Siblings remain independently cancellable and keep deterministic seeds, the source prompt, and durable batch position through long-video chains. Every print is admitted through one durable /api/generation-batches operation, chunked at the machine's advertised limit; held children survive app/server restarts with an error and retry action. A machine that cannot carry a request refuses it inline by name, and nothing is queued. A partial result names each failed variation and reviewed prompt, plus any separate unconfirmed-cancellation caveat, while keeping successful prints. Library shows Batch N of M and the source prompt when that provenance is present.

When iOS suspends the app, interrupted generation streams carry a structured recovery marker rather than relying on localized WebKit error text. On resume, Mold checks the frozen host and bounds any pre-ID queue or durable-chain join to the original submission window, so a later fixed-seed duplicate is never cancelled or reported as the interrupted print.

Library

Library merges prints from every saved host, newest first. Unavailable hosts are reported without hiding media from hosts that did respond.

Pinch the grid with two fingers to resize the thumbnails, between two across (largest) and five across (smallest). This is the iPhone equivalent of the thumbnail-size slider in the web and desktop Library. The size is remembered on the device and starts at three across; it is stored separately from the desktop and web setting, so resizing on your phone never changes the grid on your Mac. One-finger scrolling through the grid is unaffected.

Press and hold an image to open the native iOS image menu for Share, Save to Photos, Copy, Copy Subject, and Look Up. That menu also offers Upscale… for images and Framewise upscale… for videos. Tap Select to select multiple prints; with one image or video selected, the action bar offers the same upscale flow, and it can also select all loaded prints, clear the selection, or delete the selection. Deletion removes every matching copy from each reachable saved host, including legacy copies whose auto-save filename differs. If a host cannot complete the delete, its copy remains visible and Library reports the partial cleanup.

Tap a tile to open the full-screen viewer:

• images are shown uncropped;
• generated images open this same viewer when tapped;
• videos stream from their owning host with native playback and seeking;
• Upscale… enlarges an image, while Framewise upscale… queues a durable video job with pause, resume, and cancel; missing Real-ESRGAN weights download on first use and temporal flicker may remain;
• use Save video to add the original MP4 to Photos;
• swipe left or right to move through the loaded prints;
• use Copy image or Save photo instead of the system long-press menu;
• use Use as prompt to restore recorded generation settings, or Use as source to attach a still to the next compatible generation.

New installs also save completed stills to the iPhone Photos library automatically. Disable Settings → Photos → Save to Photos automatically if you want outputs to remain only in Mold Library. Post-generation upscaling saves both the original and upscaled image. Video generation offers the same Framewise upscaler picker, and the just-generated viewer can queue it after the original is published; videos remain on their Mold host.

Prints added since the prior Library visit carry a New badge. Images enlarged by post-generation or standalone upscaling carry an Upscaled badge on both iPhone and desktop.

Collections, favorites, tags, and titles

When a connected host supports Library organization, a Prints | Collections | Trash row appears under the Library header (each scope with its count; the row only appears when a host supports it). In Prints, a scrolling chip row filters the grid: ♥ Favorites, your most-used tags (the rest behind More…), and a chip per machine when several are connected. Favorite prints carry a ♥ badge on their tiles.

In Select mode the action bar gains Add to collection (a checklist with a New-collection input), Tag (add or remove tags, with suggestions from every host), and a ♥ toggle. Edits apply to every copy of the print on every reachable machine; if one machine cannot be updated, Library says so inline and keeps the rest.

Collections lists your collections merged across machines (cover, name, count, and which machines hold them) with a New collection row. The … menu renames or deletes a collection (two taps; its prints stay in the Library). Tap a collection to browse it; Select there offers Remove from collection.

Tap a print and use Info to edit it in place: the title (Done saves; blank clears it), ♥, tags, and its collections. The viewer's title line shows the print's title, or its prompt while untitled.

A print developed with an identity photo also lists that provenance in Info: the photo's filename, the first characters of its SHA-256, and the effective strength and start step. Saved metadata records the digest, never the face bytes, so Use as prompt restores the two knobs and re-attaches the photo only when this device still holds it. When it does not, Mold says so instead of rendering a different person.

Trash

On hosts with the server-side trash, deleting moves prints to the Trash scope instead of erasing them. The Trash grid shows how long each machine keeps trashed prints (change it from the host's detail screen), and each tile counts down to its purge ("Purges in N d"). Select offers Restore and a two-step Delete forever; Empty trash in the header purges everything after a confirming second tap. Machines without a trash keep the old immediate delete, and Mold's wording never promises recoverability there.

Reusing settings switches to the print's host and restores the model when it is installed there. If the original model is unavailable, Mold clearly identifies the compatible fallback and removes non-portable adapter/component choices. On a print made scene by scene by an older build, Use as prompt restores a plain one-shot clip built from the first scene's prompt; the per-scene provenance stays on the print.

Authenticated hosts issue a short-lived read-only media ticket for the selected file. The app never puts a long-lived API key in a video URL or buffers an entire video into phone memory.

Models

Models combines installed models with live Hugging Face and Civitai results. You can search, filter All/Images/Video, filter by source, family, or model kind (Models, LoRAs, CLIP, text encoders, VAEs, tokenizers, ControlNet), sort by downloads, rating, or recency, inspect download contents and installed components, and include NSFW entries explicitly. Every card identifies the model kind; mature results carry an 18+ NSFW badge. The detail sheet repeats those classifications, uses a kind-specific weights label, and shows available description, source, license, tags, format, and popularity metadata. When a manifest model has multiple quantizations, its detail sheet shows variant chips with checkpoint sizes; the selected chip is the exact model the Pull action downloads.

The family list comes from the browsed host. If that host is unreachable when Models opens (or its saved key has not been read from the Keychain yet) the list falls back to every family Mold has seen this session, from both catalog results and the installed models on your hosts. That set only grows, so choosing a family never shrinks the list to that one choice, and it can offer families that are not on the current page. The real list, and any search that failed, reload as soon as the host answers.

The host selector controls where Models browses. Pulling a model can target a different ready host without changing the host selected in Create. With more than one target available, Mold opens a host picker.

Pull buttons reflect the server state immediately:

• Connecting... while the download event stream opens
• Starting... while the request is accepted
• Queued while waiting for a download slot
• Pulling N% during transfer

Active downloads stay pinned above results and can be cancelled. Installed model details offer Load on GPU, Unload from GPU, and a guarded Remove from host action.

Host details

Tap a saved host to inspect its current state. The detail screen shows:

• GPU/VRAM, CPU, and RAM telemetry;
• every GPU's utilization, VRAM, lifecycle state, and enable/disable control;
• free and used storage for the models filesystem;
• queued/running generation work, with a confirmed Cancel action for queued work and active singleton generations on current hosts, and loaded models;
• queued/active model downloads with progress;
• the models installed on that host.

Hosts with the server-side Library trash add a Library card: choose how long that machine keeps trashed prints (Trash retention, 0 keeps them forever; a value pinned by an environment variable is read-only and names the variable) and see Prints in trash: N with a two-step Empty trash. The setting lives on the host itself; it applies to every app that talks to it.

From the same screen you can rename or retry the host, select it for Create, unload a model, open it in Models, or forget it. Forgetting a host also deletes its API key from the iOS Keychain. Queue cancellation uses that host's Keychain-authenticated route and refreshes the list after the server confirms it. Running inference is revoked cooperatively at its next safe point.

Settings and themes

Open Settings from the sliders button in the header. Mobile settings currently cover:

• Look: the five Mold Studio themes — Mocha, Safelight, Blueprint, Graphite, Nebula — each in a light and a dark tone
• Light or dark: Match phone, Light or Dark. Match phone follows iOS between the chosen theme's own two tones; it never swaps the theme itself
• Photos: automatically save newly generated stills to the iPhone photo library (on by default)
• Model licenses: review and accept the selected host's model licenses
• Library: Tag new prints with their title; offer a titled print its own title slug as a tag in Create (on by default)
• Remote hosts: saved-host count and a shortcut to manage them
• Compute devices: enable or disable each GPU on the selected host; a busy GPU finishes its current stage before disabling
• About: app version, remote-only processing, TestFlight updates, and equal project-owner credit for core contributors James Brink and Jeffrey Dilley, plus an external link to the Mold privacy policy

Fresh installs start on the Safelight theme with Match phone off, and with Photos auto-save and title tagging enabled. Existing users keep any valid saved choices, with auto-save and title tagging enabled when upgrading from a settings record that predates either option.

The appearance choice updates both the WebView and native iOS system chrome so status-bar content remains readable. Themes change the app chrome but never recolor generated photos or videos.

The app intentionally prevents WebKit input-focus and double-tap page magnification and suppresses rubber-band scrolling. The Library's horizontal swipe navigation and its grid pinch-to-resize gesture remain enabled. iOS system-level accessibility Zoom is separate from WebView page zoom.

Updates and current boundaries

Internal TestFlight builds are created automatically after mobile-relevant changes pass the iOS workflow on main. The pipeline validates the bundled mobile index.html, Mold icon catalog, native archive, App Store Connect processing, and internal tester access before it is considered complete.

The iPhone app focuses on remote Make, Queue, Images, Styles, Machines, and appearance settings. Video is a section inside Make, not a separate screen: pick a clip style, describe the shot, choose the duration, and generate while Mold keeps progress and cancellation attached to the exact host. Ask for a clip longer than one render and the machine splits, carries, and stitches the work for you — it arrives as one video in one queue row. An optional opening image, with its source strength and fit controls, is a disclosure in the primary Create stack beside the other source media rather than inside the Advanced sheet; the primary ↺ Reset clears it while Reset advanced leaves it alone. Use desktop or the CLI for a local engine, scene-by-scene scripted sequences and the full jobs-administration workspace, RunPod provisioning, engine configuration, and desktop Stable/Nightly self-update controls.

For server networking and deployment options, continue with Remote Workflows. For supported model-family controls, see Feature Support.

Troubleshooting

Discovery finds nothing

• Confirm Local Network access is enabled for Mold in iOS Settings.
• Confirm the phone and host are on the same LAN; Bonjour does not browse a remote tailnet.
• Confirm the server was not started with --no-mdns or MOLD_MDNS=0.
• Add the IP address or Tailscale MagicDNS name manually.

The host test fails

• Open http://host:7680/health from another device on the same network.
• Confirm the server listens on 0.0.0.0, not only 127.0.0.1.
• Check the firewall, Tailscale ACL, port, scheme, and API key.

A video will not play

Update the remote Mold host. Current authenticated video playback needs the short-lived gallery media-ticket endpoint so iOS can make native Range requests without exposing the API key.

Mobile redesign foundation

The phone uses Make, Queue, Images, Styles, and Machines. Make shares the desktop’s Still picture, Short clip, and 3-D object sections and remembers the style last used in each. Prompt and required photos stay on the main screen; Detail, guidance, seed, mesh controls and other secondary controls live in More settings. Shape and Length remain in the main flow. Name and organize is optional. Generate stays above the software keyboard.

Queue combines this phone’s work with live work from connected machines. Being made, Waiting, and Needs attention use the existing lifecycle and wait contracts. Opening details preserves the draft; restoring settings is explicit. Full queue records remain in memory. Finished shows at most 20 recent jobs known to this phone, not a durable fleet history; saved results remain in My images. Offline work is labeled as last known and cannot be changed until its machine reconnects.

Saved machines appear before Add a machine, which contains pairing, nearby discovery and manual address entry. The connection controls start expanded on first setup. This foundation retains the existing Images and Styles workflows; further visual refinements and physical-device acceptance remain tracked in issue #1628. Scene authoring is intentionally absent.