Terminal UI

mold includes an interactive terminal UI for browsing models, tuning parameters, generating images with live progress, and previewing results — all without leaving the terminal.

Built on ratatui with Kitty graphics protocol support for pixel-perfect image preview in terminals like Ghostty, kitty, and WezTerm. Falls back to halfblock rendering in other terminals.

Beta

The TUI is under active development. Core generation, script-mode video authoring, model management, gallery, settings, theming, and image preview work well.

Quick Start

mold tui

The TUI launches in the Create workspace with your cursor in the prompt field. Type a prompt, press Enter, and watch the progress panel as your image generates. The result appears in the Preview panel and is saved to ~/.mold/output/.

TIP

The tui feature must be compiled in. Pre-built releases and the Nix package include it by default. If building from source, add --features tui to your build command.

Auto-Start Server

By default, mold tui automatically starts a background mold serve process if no server is already running. This keeps models hot between generations for faster subsequent runs. The server is killed when you quit the TUI.

• mold tui — auto-starts server on localhost:7680
• mold tui --local — skip server, use local GPU only
• mold tui --host http://gpu:7680 — connect to an existing remote server

Server Logs

When the TUI auto-starts a background server, logs are written to ~/.mold/logs/ with daily rotation. This is useful for debugging generation failures since the server's stderr is suppressed while the TUI controls the terminal.

Log files follow the naming pattern mold-server.YYYY-MM-DD.log and are automatically cleaned up after 7 days (configurable via logging.max_days in ~/.mold/config.toml).

To view live logs while the TUI is running, open a second terminal:

tail -f ~/.mold/logs/mold-server.$(date +%Y-%m-%d).log

You can also enable file logging for manual mold serve with --log-file, or permanently via the config file:

[logging]
level = "info"
file = true
# dir = "~/.mold/logs"
# max_days = 7

Workspaces

The TUI shares the five Mold Studio workspaces with the desktop, web, and iPhone apps, shown as tabs at the top of the screen:

Workspace	Key	Purpose
Create	1	Write prompts, tune parameters, generate images/video
Library	2	Browse prints from this machine and every known host
Models	3	View installed and available models
Machines	4	Connect remote Mold hosts, telemetry, queue, target
Settings	5	Theme picker plus file-backed and DB-backed settings

Switch workspaces with Esc then 1–5, or click the tabs. Alt+1–Alt+5 works from anywhere, and Esc from any other workspace returns to Create. The chain composer opens with c from Create — it is a Create sub-mode, not a tab.

Command Palette

Ctrl+K opens the command palette from any workspace or focus — the TUI's version of the GUI surfaces' ⌘K launcher. Type to filter, Up/ Down to select, Enter to run, Esc to close. It covers navigation (all five workspaces, the chain composer), actions (toggle Advanced, randomize seed, expand prompt, prompt history, help, quit), and switching between all eleven theme presets.

Create View

The main workspace with four panels:

• Prompt — Multi-line text area (Shift+Enter for newlines, emacs keybindings). Required, except for an LTX-2 / LTX-Video model that already has a source image attached, where an empty prompt is accepted and prompt expansion is skipped for that run
• Parameters — six essentials rows plus the Advanced accordion
• Preview — idle hint; while generating, the latest live latent preview frame (for families that stream denoise previews — FLUX.1, Flux.2, Z-Image, and Wan 2.1/2.2) with the denoise progress line beneath it; then the finished print with a model · seed · time · host caption (Kitty/sixel/halfblock rendering). MOLD_STEP_PREVIEW=0 disables server preview streaming
• Timeline — the glyph-styled session log (• info, ✓ done with stage timings, ! warning, ✗ error, ★ model loaded), including a ✓ Saved <file> entry per print; shows "— idle. no runs this session." until the first run. Hidden when tui.show_timeline is off or the terminal is too short.

The model's description is the dim line under the Model row; host and memory telemetry live in the Machines workspace and the chrome host chip.

Essentials

Row	Shows	◀▶ / +/-	Enter
Model	human-readable model name + description	—	fuzzy model selector
Size	1024 × 1024	cycle aspect presets	type an exact WxH
Detail	●●●●○○○○ 28 step dots	adjust steps	—
Prompt strength	guidance	adjust	—
Seed	random / fixed · 42	cycle seed mode	type an exact seed
Batch	image count	adjust	—

Size's ◀▶ cycles 1:1, 3:2, 2:3, 16:9, and 9:16 presets fitted to the model's default pixel area (64-aligned). Seed modes are random (new seed every run), fixed (reproducible), and increment (+1 per run) — Ctrl+R still cycles them from anywhere.

Advanced accordion

Press A anywhere in Create (or Enter/→ on the ▸ Advanced row) to open the disclosure; Enter or → on a section expands it, collapsing any other, and ← collapses. Each collapsed section shows a summary (default, off, none, png, or the set value) and the header carries a count of advanced values that differ from their defaults. The open state and expanded section persist across sessions (tui.advanced_open / tui.advanced_section).

Section	Rows
Scheduler & sampling	Scheduler (CFG models), Expand prompt, Offload
Negative prompt	inline editor (CFG models; Alt+N jumps here). Wan models prefill their tuned default — leave it for the default, edit to replace it, clear it to send an explicit empty negative
Source image	Source, Strength, Mask, ControlNet (per model)
Identity photo	Photo, Strength, Start step (only for a checkpoint the server advertises as identity-capable)
LoRA	LoRA path + scale
Upscale after generate	post-generate upscaler (Enter picks, (off) clears)
Output format	png / jpeg / gif / apng / webp / mp4
Video	Frames, FPS; Wan Flow shift; LTX-2 Pipeline, Audio default/on/off, Spatial native/1.5×/2×, Temporal native/2×, STG scale/blocks, CFG rescale, Modality scale, Guidance skip
File under	Title, Tags, Collection — where the print lands in the Library

The LTX-2 Pipeline row cycles through Auto, one-stage, two-stage, two-stage-hq, and distilled. Auto omits the request field so the server keeps checkpoint selection authority; each explicit recipe selects MP4 and is shown in the collapsed Video summary and Advanced badge. Guidance and the Negative prompt section follow the selected recipe's CFG contract. Pipelines that require an audio file, source video, keyframes, a retake window, or an IC-LoRA—and the audio-only t2a output—remain intentionally absent until the TUI can author and handle those inputs and outputs end to end.

For Wan models the Video section instead exposes Flow shift (the family's primary quality/character knob, request field sample_shift); it is absent until touched and never appears for other families. LTX-2's optional STG scale/blocks, CFG rescale, modality scale, and guidance-skip rows keep their request fields absent until edited, so pipeline constants stay authoritative.

Identity photo (PuLID)

The Identity photo section appears only while the selected checkpoint's /api/models[] entry advertises supports_identity — face-identity conditioning is qualified for a fixed list of FLUX and SDXL checkpoints (see the Identity Photos guide), on a server built with PuLID support. It is never inferred from the family or from how your local mold was compiled, because the render happens on the server.

Enter on the Photo row opens a path picker. The file is opened without following symlinks and bounds-checked (PNG or JPEG, at most 16 MiB encoded, 8192 px per axis, 32 MP) before it is accepted, so a bad path is refused in the picker rather than after your job has taken a queue slot. Leaving the field empty clears the photo. Strength moves in 0.1 steps across 0.0–3.0 (default 1.0) and Start step across 0–steps-1 (default 0), so the form cannot express a value the server would reject; lowering the step count pulls Start step back with it.

Switching to a model that cannot take the photo keeps it — the photo is your choice, and switching back must not have lost it — but shows the server's refusal on the Photo row and blocks Generate. That is deliberate: rendering silently without the face would produce a print that looks fine and is simply not the person you asked for. ↺ Reset to model defaults clears the photo and both knobs.

Prints that carried a face reference show Identity in Library Details and in the full print view: the photo's name, a short digest, the strength, and the start step. The face bytes themselves are never recorded.

The Audio row is capability-driven: a checkpoint that advertises missing audio assets does not show it. default leaves the field absent so the server's pipeline constant remains authoritative; on and off send an explicit choice. Switching Audio on also selects MP4, the only container that can carry LTX-2's synchronized audio.

The Spatial and Temporal rows use the same optional latent-upscale fields as the other generation surfaces. native omits the matching request field; explicit choices are summarized on the collapsed Video row and reset when the model changes to a family that cannot use them.

File under

The last section files the print as you make it, so it arrives in the Library already named and organized instead of needing a second pass through the gallery. It has no capability gate — every model produces a print.

Row	What it takes
Title	The print's name (≤120 characters). Blank clears it.
Tags	Comma-separated, up to 20 of 1–64 characters. Blank clears them.
Collection	One collection by name; created on the host when it does not exist. Blank clears.

Enter on a row opens its editor. Entry is checked before the editor will close: an over-long title, a tag with a control character, or a collection name with no letters or digits keeps the popup open with the reason beneath what you typed, so nothing the host would refuse reaches a request. Esc abandons the edit. Tags are trimmed and de-duplicated case-insensitively, and a collection name's whitespace is collapsed so the same name means the same collection on every machine.

All three are absent until you touch them — an untouched form sends exactly the request it always did.

While Tag by title is on (Settings ▸ Library, on by default), a titled print also picks up its own title slug as a tag. That is a client decision, not the host's, so the TUI shows it before you generate: the Tags row reads auto: smurf-village and the collapsed section summary repeats it. Turn the setting off and the derived tag disappears; the tags you typed stay.

 ▾ File under  “Smurf Village” · 2 tags · in Blue Period · auto: smurf-village
  Title         Smurf Village
  Tags          village, blue · auto: smurf-village
  Collection    Blue Period

Filing is per-print intent rather than a preference: it is not restored on the next launch, it is not saved per model, and only ↺ Reset to model defaults clears it. Generating keeps it, so a batch's siblings share the filing.

If a host cannot apply what you filed — it is running without a metadata database, or the collection went away between listing it and pressing Generate — the print is still rendered and saved, and the TUI says what was dropped: a one-line ! advisory on the Create view (several are joined with ·) plus an entry per advisory in the Timeline. It is an advisory rather than an error because the render succeeded, and starting the next generation clears the line. The same row carries any other adjustment a host reports about an accepted request, such as a lip-dub clip retimed to its reference.

A titled print's file is named mold-{model}-{ts}[-{idx}]~{title-slug}.{ext}, the same shape the server's gallery writes — renaming the print later never renames the file.

The ↺ Reset to model defaults action row at the bottom restores every parameter (keeping the model and your prompt). qwen-image-edit shows a source image and negative prompt without img2img strength or mask controls.

Model Selector

Press Enter on the Model field or Ctrl+M from anywhere:

• Type to fuzzy-filter the model list
• j/k, arrow keys, or scroll wheel to navigate
• Enter to select — parameters update to model defaults
• Esc to cancel

Prompt History

Previous prompts persist across sessions in ~/.mold/prompt-history.jsonl:

• Up/Down arrows at top/bottom of prompt recall history
• Ctrl+P/Ctrl+N also navigate history
• / in navigation mode opens fuzzy search over all prompts

Shell Keybindings

The prompt editor supports standard emacs/shell keybindings:

Key	Action
Ctrl+A	Beginning of line
Ctrl+E	End of line
Ctrl+U	Kill to start
Ctrl+W	Delete word back
Ctrl+D	Delete forward
Ctrl+F	Forward char
Ctrl+B	Backward char

Library View

The Library merges prints from every machine you know about into one grid — the local output directory (~/.mold/output/ or MOLD_OUTPUT_DIR), the connected server, and every host registered in the Machines workspace (fetched with each host's saved API key). Cross-host copies of one print are collapsed by filename — the same identity rule the desktop's unified gallery uses — with the local copy preferred when one exists. The header is honest about sources: 28 prints for local-only, 28 prints · hal9000 when everything came from one remote machine, 28 prints · all machines once prints span more than one, and a dim · 1 host offline suffix when a host didn't answer the scan. Offline hosts never break the merge — their prints just drop out until the next rescan (entering the Library rescans automatically when the last scan is stale or a host was added/removed). Only live prints are listed — a print moved to the trash keeps its row and its bytes but leaves the Library until restored from a Trash view.

On wide terminals a Details side panel shows the selected print: its thumbnail, the print's title when it has one (bold, above the prompt), the wrapped prompt (plus a dim neg: line), and Model / Seed / Size / optional video Pipeline / Machine rows — Pipeline is the recipe the engine actually ran, including an Auto-selected LTX-2 recipe, while Machine names the host the print lives on ("This Mac", a host name, or 2 machines when copies exist on several). The panel hides automatically on narrow terminals. Full detail mode shows the title and the same runtime pipeline when recorded.

Grid Mode

Key	Action
h/j/k/l	Navigate the grid
Arrow keys	Navigate the grid
Enter	Open detail view
e or r	Recall into Create (edit)
u	Upscale with AI model (runs on the owning host)
d	Move print to the trash (with confirmation)
o	Open in system viewer
/	Filter by prompt, model, or filename
Esc	Clear the filter, then back to Create

Typing after / filters live (case-insensitive, matching prompt, model name, and filename); Enter keeps the filter applied, Esc clears it. d moves a print to the trash on every machine that holds it — the confirmation names the machine count before anything happens. The trash wording appears only when every owning machine can actually trash: the local move needs the metadata DB (the bytes land in <output_dir>/.trash/ beside a tombstone, recoverable from any Trash view), and a server must advertise gallery.trash in its capabilities. Otherwise the hint and the confirmation fall back to honest permanent-delete wording — an older server or a DB-less local scan really does delete. Recall, upscale, and removal work on remote prints exactly like local ones; requests route to the machine that owns the print.

Detail Mode

Press Enter on a grid thumbnail to see the full image with all metadata.

Key	Action
e	Load into Create (edit)
r	Regenerate immediately
u	Upscale with AI model
d	Move to trash
o	Open in system viewer
j/k	Previous / next image
Esc	Back to grid

Thumbnails

Thumbnails are cached at ~/.mold/cache/thumbnails/ and generated automatically on first scan and after each generation. Delete the cache directory to force regeneration.

Chain Composer

The chain composer authors mold.chain.v1 TOML for multi-clip video chains — LTX-2, LTX-Video, and Wan 2.1/2.2. Frame counts validate on the selected family's own grid (8n+1 for LTX-2, 4n+1 for Wan) and the seam carryover is per family: LTX-2 carries a 17-frame motion tail, Wan's image-conditioned checkpoints continue from a single seed frame, and text-to-video checkpoints join independent clips. Press c from Create's navigation mode to open it (Esc returns to composing — a chain in progress survives switching workspaces). It lets you build per-stage prompts, frame counts, source images, and smooth / cut / fade transitions, then submit the normalised script through the same chain endpoint used by mold run --script.

Models View

See all installed and available models with family, size, defaults, and status.

Key	Action
j/k	Navigate the model list
Enter	Set as default model
p	Pull (download) a model
u	Unload the active model (GPU)
Esc	Back to Create

Machines View

Manage every Mold server the TUI can generate on — the same multi-host Machines workspace as the desktop, web, and iPhone apps. The left pane lists machines with a status dot (green ready, yellow connecting, red offline), the machine name, and a dim hardware line (GPU, VRAM, backend, host:port); This Mac (or This machine) is always the first row. The right pane shows the selected machine's telemetry (GPU, VRAM, models-disk storage, loaded models, uptime, version) and its live queue lanes — ▶ running jobs with elapsed time and GPU ordinal, ● queued jobs with their position. Offline hosts stay listed and recover automatically when the server comes back. When a host has several GPUs, the detail pane lists every device with its stable ID, utilization, VRAM, loaded model, active work, and lifecycle state.

Key	Action
j/k	Select a machine (or a queue lane with detail focus)
Enter	Set as the generation target (again reverts to Auto)
Tab	Switch focus between the host list and the detail lanes
c	Connect a machine (also in the ⌘K palette from anywhere)
d	Forget the selected host (confirms; deletes its API key)
g	Select the next GPU on the current machine
e	Enable or disable the selected GPU
r	Refresh telemetry and queue now
[ / ]	Select the previous or next GPU/MIG device
e	Enable or disable the selected device (when advertised)
x	Cancel the selected queued job (detail focus, confirms)
Esc	Back to Create

Disabling a busy GPU removes it from future scheduling immediately but lets its current stage finish before the owner thread exits. Enabling it starts a fresh owner thread. If every GPU is disabled, the server remains available for settings, telemetry, downloads, and other maintenance work. Live disable is hidden and ignored unless the selected host advertises both devices.lifecycle and authoritative V2 dispatch. A persistently-disabled, startup-selected device may still offer Enable on restart when the host advertises devices.restart_enable.

Connecting a machine

Press c for the stepped connect flow: enter the server address (bare hostname, host:port, or full URL — bare names default to port 7680), then an optional API key (masked; Enter skips it), and the TUI tests the connection. On success the host is saved with its display name taken from the server's hostname; if the same server is already registered (matched by its instance ID, URL, or id) you get "Already connected as …" instead of a duplicate row. On failure the error is shown with Enter to retry or e to edit the address.

Hosts persist in the settings DB (mold.db) under tui.hosts.v1, with each host's API key stored in its own tui.host_key.<id> settings row — keys are sent as the x-api-key header, never placed in URLs, and are deleted when the host is forgotten.

Generation target

Enter on a machine row makes it the sticky generation target (persisted as tui.generate_target): This Mac forces the local engine, a remote host routes every Generate to that server with its API key — and never silently falls back to local. If a targeted host is unreachable the run fails with an error naming the host so you can fix it in Machines or press Enter on the target row again to return to Auto (remote when connected, local fallback — the default).

Settings View

Edit settings without leaving the TUI. Bootstrap values such as paths, ports, credentials, and logging persist to config.toml; user preferences and per-model generation defaults persist to the SQLite settings DB at $MOLD_HOME/mold.db.

Key	Action
j/k	Navigate settings
+/- or L/R	Adjust numeric or cycle toggle values
Enter	Edit text/path field (opens popup) or toggle
Tab	Switch between Appearance and Configuration
Esc	Back to Create

Appearance

The Appearance panel renders the eleven theme presets as bordered cards — three swatch dots showing each preset's background, accent, and info hues, the preset name, and a short palette descriptor. The selected card carries the focus-colored border, and the panel header shows the active slug as theme · <slug>.

Arrow keys move the selection in two dimensions: Up/Down move by grid rows (Down past the bottom row drops into the Configuration list), while Left/Right and +/- cycle linearly through every preset. Every move applies the theme immediately and persists it under tui.theme. On short terminals the grid scrolls by whole card rows to keep the selection visible.

Preferences

The Configuration list starts with a DB-backed Preferences section. Each toggle persists to mold.db the moment it flips.

Row	Key	Default	Effect
Format	tui.default_format	png	Seeds a fresh session's Format parameter (a saved session or per-model preference still wins)
Reduce Motion	tui.reduce_motion	off	Disables TUI motion effects (consumed by upcoming releases)
Show Timeline	tui.show_timeline	on	Shows the Timeline on the Create view (consumed by upcoming releases)
Confirmations	tui.confirm_destructive	on	When off, destructive actions — deleting a print, removing a model, deleting a chain stage — run immediately without a confirmation popup

A Library section follows: Trash (days) edits the shared gallery.trash_retention_days key — how long trashed prints survive before the server's retention sweeper purges them (0 keeps them forever, max 3650, default 30). The value persists through the same settings-DB surface mold config set and the server's config API use, so every surface reads one window; MOLD_GALLERY_TRASH_RETENTION_DAYS overrides it with the usual (env) indicator.

Tag by title sits beside it and edits generate.auto_tag_title (on by default): whether a titled print also picks up its own title slug as a tag. That is a client decision — the server never auto-tags, because it cannot tell a typed title from a scripted one — so the same key drives mold run and Create ▸ Advanced ▸ File under, which discloses the derived tag before you generate. Turning it off changes nothing about prints you already made.

Field Types

• Numeric (port, width, steps, etc.) — adjust with +/- keys
• Boolean (metadata, expand enabled, etc.) — toggle with Enter or +/-
• Toggle (T5 variant, log level, scheduler) — cycle with +/- or Enter
• Text/Path (model name, directories, prompts) — Enter opens edit popup
• Read-only (model file paths) — displayed dimmed, not editable

Environment variable overrides are shown with an (env) indicator in yellow. Per-model defaults show resolved values from the manifest (not raw config None values), so you always see the effective runtime value.

Model Defaults

The Model Defaults section shows settings for a specific model. Use Left/Right on the Model selector row to cycle through configured models. Editable fields include steps, guidance, dimensions, scheduler, negative prompt, LoRA path, and LoRA scale. File paths (transformer, VAE) are read-only since they are managed by mold pull.

Qwen-Image-Edit

The TUI treats qwen-image-edit as a distinct edit family:

• single source image only in the TUI
• no img2img strength
• no inpainting mask
• no ControlNet controls
• LoRA controls are available because qwen-image-edit is LoRA-capable
• default width/height derived from the selected source image at roughly 1024x1024 area

Local inference uses the Qwen2.5-VL multimodal edit encoder. In v1 the TUI keeps the flow single-image only, even though the CLI and API accept multiple ordered --image inputs for qwen-image-edit.

Navigation

Press Esc to enter navigation mode, where number keys and arrows switch views.

Global Shortcuts

Key	Action
Esc	Unfocus / navigation mode
1 – 5	Switch workspace (in navigation mode)
Left / Right	Cycle workspaces (in navigation mode)
Alt+1 – Alt+5	Switch workspace (from anywhere)
Tab	Cycle focus to next panel
Shift+Tab	Cycle focus to previous panel
Ctrl+C	Quit
q	Quit (when not in a text field)
?	Show help overlay

Create Shortcuts

Key	Context	Action
Enter	Prompt	Start generation
Enter	Parameters	Activate row (selector/toggle/expand)
Ctrl+G	Any	Start generation
Ctrl+M	Any	Open model selector
Ctrl+R	Any	Cycle seed mode
Ctrl+E	Any	Expand prompt via LLM
Alt+N	Any	Open Advanced → Negative → focus editor
c	Navigation	Open chain composer
A	Anywhere	Toggle the Advanced accordion
+/-	Parameters	Adjust value / expand-collapse a section
j/k	Parameters	Navigate rows (flat, sections included)

Mouse Support

• Click tabs to switch views
• Click panels to focus them
• Click parameter rows to select and activate
• Click gallery thumbnails to select, double-click for detail view
• Click model rows to select
• Scroll wheel navigates lists and popups

Session Persistence

All settings are saved to ~/.mold/tui-session.json after each generation and restored on next launch:

• Prompt and negative prompt text
• Model selection
• All generation parameters (dimensions, steps, guidance, seed mode, batch, format, scheduler, lora, expand, offload, strength)

Use the ↺ Reset to model defaults row at the bottom of Parameters to restore model defaults without losing your prompt. Unloading the active model moved to the Models workspace (u).

Generated images are saved to ~/.mold/output/ by default (override with MOLD_OUTPUT_DIR env var or output_dir in config). All images include embedded PNG metadata that preserves the full generation parameters, making them portable across machines.

Routing

Where Generate runs is owned by the Machines workspace, not the Create form: press Enter on a machine row to pin it as the sticky generation target (persisted as tui.generate_target), or leave it on Auto — try the connected server first, fall back to the local GPU. mold tui --local pins the session to the local engine. A pinned remote host never silently falls back to local; an unreachable target fails with an error naming the host. See the Machines section above for connecting hosts and API keys.

Image Preview

The TUI auto-detects your terminal's graphics protocol at startup:

Protocol	Terminals	Quality
Kitty	Ghostty, kitty, WezTerm	Pixel-perfect
Sixel	foot, xterm, mlterm	Full color
iTerm2	iTerm2, Hyper	Full color
Halfblocks	Everything else	Unicode blocks

Building with TUI Support

The TUI is behind the tui feature flag on mold-ai:

macOS (Metal)

cargo build --release -p mold-ai --features metal,tui

Linux (CUDA)

cargo build --release -p mold-ai --features cuda,tui

All features

cargo build --release -p mold-ai --features metal,preview,discord,expand,tui

The Nix flake, pre-built releases, and Docker images include the TUI by default.

Motion

Switching workspaces fades the new content in and a finished print sweeps its caption row — the TUI's only two animations, matching the design system's restraint budget. Turn them off with the Reduce motion preference in Settings (persisted as tui.reduce_motion) or the MOLD_TUI_NO_MOTION=1 environment variable (useful for scripted terminal captures).

Theme

The TUI ships eleven theme presets, selectable live from Settings → Appearance — a grid of theme cards navigated with the arrow keys (Left/Right also cycle linearly), applying immediately; the choice persists in the metadata DB under tui.theme.

The default is Studio Dark — the same Mold Studio look as the desktop, web, and iPhone apps — with its dual-accent role model: the warm safelight accent marks focus, selection, and primary actions, while the cool halide accent marks info and live state.

Preset	Slug	Notes
Studio Dark	studio-dark	Default. Mold family, dual accent
Studio Light	studio-light	Mold family, dual accent
Safelight Dark	safelight-dark	Warm darkroom family, dual accent
Safelight Light	safelight-light	Warm darkroom family, dual accent
Mocha	mocha	Catppuccin Mocha — the pre-Studio default
Latte	latte	Catppuccin Latte (light)
Ristretto	ristretto	Monokai Ristretto
Gruvbox	gruvbox	Gruvbox Dark (hard)
Tokyo	tokyo	Tokyo Night (storm)
Nord	nord	Nord
Dracula	dracula	Dracula

studio and safelight are accepted as slug aliases for the dark variants.