Operator's
Manual

What it does

Canvas builds a single considered frame from sidebar instruments. Playground is an infinite board with a floating prompt bar, the whole model library, workflows and mini-apps. Workbench is a node graph you wire once and re-run. Cinema plans a timed production scene by scene. Board collects research. Midjourney Studio, Studio Apps (Higgsfield) and WebUI (open and local models) run their own engine families. Double-click the main Canvas to open Playground; the upper toolbar then switches between surfaces and Exit returns to Canvas.

What to expect

All surfaces write into the same gallery and obey the same credit and sync rules. Switching modes never loses work already generated.

Use it well

Explore in Playground at low resolution and low count, finish in Canvas at full resolution, and wire anything you have done three times into Workbench.

What it does

Open Studio in Canvas mode, add a sketch or reference in the Inputs tab (optional), describe the subject in the Prompt tab, set Resolution and Engine, then press Generate.

What to expect

A shimmer placeholder holds the slot on the canvas while the run is in flight; the result lands with the labs rail attached to it.

Use it well

Start with a prompt only. Add a sketch when the geometry matters and a reference when the finish matters — then you know which input is doing the work.

What it does

Cost is computed server-side from model, count, resolution and duration before anything is sent upstream.

What to expect

The number the interface shows is a preview of the server figure, not the authority. A run that fails without producing a result refunds automatically.

Use it well

Treat the preview cost as a budget check, then read the account usage breakdown for the truth.

What it does

Upload a drawing, set its Influence, and enable any of three toggles: Geometry, Contour, Detail. Each toggle decides which part of the drawing the engine is asked to respect.

What to expect

With Detail on, linework and hatching are read as surface texture. With only Geometry on, the drawing guides proportion and the render invents its own finish.

Use it well

Start around 50% influence with Geometry and Contour on. Raise influence only if the render drifts off the package you drew.

What it does

A second image with its own influence and three toggles: Material, Lighting, Composition. It borrows finish and light without borrowing shape.

What to expect

With Composition off, the reference contributes look only. With Composition on, the frame starts arranging itself like the reference.

Use it well

Material + Lighting on, Composition off, is the default for CMF work. Add Composition only when you are deliberately copying a layout.

What it does

Free text with Save, Enhance and Randomize; a guided Prompt Builder as an alternative to free text, including a Text / No Text switch; a Negative Prompt field; and a Seed you can lock or roll.

What to expect

Enhance rewrites your sentence into a longer directive. Locking the seed makes repeat runs comparable so a single changed instrument is readable.

Use it well

Once a frame is close, lock the seed and move exactly one instrument at a time.

What it does

Resolution is Auto / 1K / 2K / 4K. Engine picks the model family, badged by quality class (Ultra / High / Fast). The canvas toolbar sets count at 1, 2 or 4 variants per run.

What to expect

Higher resolution classes and higher engine tiers are plan-gated; the badge shows the active class. Cost scales with resolution and count.

Use it well

Search at 1K and count 4, then re-run the winning seed at 2K or 4K with count 1.

What it does

Input history, Size and View, Count, Rig (camera instruments), Compose (director shot) and Regenerate, all attached to the frame you are looking at.

What to expect

Rig and Compose are popovers over the canvas; their compiled clause is appended to the prompt when you generate.

Use it well

Use View for the standard automotive/product angles and reserve Rig for lens and light nuance.

What it does

Orbit picks a view node (front, front 3/4, side, rear 3/4, rear, or multiple). The lens dial drags focal length across 14–300mm and aperture across ƒ/1.2–ƒ/16, with a live field-of-view wedge and iris.

What to expect

The rig chip reads back as `50 MM · ƒ/2.8 · AZ 135° EL 45°`. Untouched instruments stay null and contribute nothing to the prompt.

Use it well

Long lens plus wide aperture for hero product separation; 24–35mm for context and environment.

What it does

Azimuth and elevation for the key light, compiled into a directional clause — front-left, rim, overhead, near the subject line.

What to expect

Low elevation throws long shadows; high elevation flattens them. Rear azimuth reads as a rim light.

Use it well

Set light after the view is decided, so you are lighting a known angle rather than guessing.

What it does

Five instrument cards plus a searchable preset browser and a compiled shot brief showing the exact sentence that will be sent.

What to expect

Presets replace each other rather than stacking — selecting a second preset swaps the shot sentence instead of piling both into the prompt. On desktop it opens as a wide horizontal panel; on mobile it becomes tabbed groups.

Use it well

Pick a preset first, then nudge individual gizmos. The brief at the bottom is the ground truth for what you are about to send.

What it does

The rig sets direction before a render. Camera Lab re-shoots an existing frame — angle, lens and framing applied to the image you already have.

What to expect

The rig changes what gets generated; the lab changes what already exists and costs a fresh generation.

Use it well

Use the rig while searching, use Camera Lab once you have a frame worth re-shooting.

What it does

Modality switch (image or video), model picker, aspect ratio, variant count (1 / 2 / 4, capped by the model), seed, a Text / No Text flag, reference chips, style palette chips and the instrument popovers.

What to expect

Runs are concurrent and land in free slots on the board; nothing blocks while a render is in flight.

Use it well

Keep one axis fixed per sweep — same prompt, four models, or same model, four seeds.

What it does

Up to four images, each tagged subject or style, each with its own influence set by dragging vertically on the thumbnail.

What to expect

A subject chip pulls identity and shape; a style chip pulls finish and light. High influence on a style chip can start dragging the subject with it.

Use it well

One subject chip, one or two style chips. Drag style influence down before you rewrite the prompt.

What it does

A full-screen browser filtered by category (all / image / video) and provider, searchable, showing the credit cost of every engine.

What to expect

Flagship models are badged and plan-locked models show a lock. Capability warnings state up front when a model ignores image references, has no end-frame support, or cannot extend a clip.

Use it well

Read the capability warning before blaming your reference — some engines simply do not accept one.

What it does

The board becomes a scrollable card feed with a bottom control dock: prompt, add image, and chips for Model, Ratio, Length, Resolution, Count and Sound.

What to expect

Parameter pickers open as full-height sheets with an explicit close mark so nothing hides under the toolbar. The instrument gizmos are fully interactive on touch.

Use it well

Use the dock chips as your checklist before every run — they show current state without opening a sheet.

What it does

Ready-made Playground workflows grouped into categories such as Cinematic Video, Motion & Camera, Effects, Upscale & Restoration, Concept Art & Style, Product Form & CMF, Character & Fashion and Backgrounds & Masking. Each declares its input slots (subject, style, start and end frame) and the engine it runs on.

What to expect

Opening a workflow fills the slots and prompt program for you; you still choose when to run. Entries marked coming soon cannot be run yet.

Use it well

Use a workflow when the task has a name (extend, relight, image to video); write your own prompt when it does not.

What it does

Sixteen one-click apps in four groups — Design & Product, Edit & Finish, Motion and Cinema — such as Angle Explorer, CMF Studio, Smart Expand, Light Studio, Animate Image, Bridge Frames and Product Film.

What to expect

The shelf offers the apps that fit what you have selected (nothing, one image, several images, video). Design, edit and motion apps open the matching Playground workflow; Cinema apps hand your selection to a new Cinema production.

Use it well

Select first, then open the shelf — the list narrows to what will actually work.

What it does

Playground workflows and mini-apps are single-step recipes run on the board. Workbench is a graph of many steps you wire and re-run.

What to expect

Both write to the same Library and use the same credit rules.

Use it well

If a recipe needs more than one generator in a row, or a fan-out to several angles, build it in Workbench.

What it does

Edit (masked local editing with before/after), Scene (background and environment), Material (finish and surface), Camera (angle, lens, framing), Form (silhouette and proportion), Render (studio, clay, technical, photographic), Style, Poster, Sketch, Upscale, UpMax and the Atelier Suite. Full productions live in Cinema.

What to expect

Each lab preserves what it is not responsible for. Material Lab keeps the geometry; Form Lab keeps the finish.

Use it well

If two things are wrong, run two labs. One lab asked to fix two things will reinvent the frame.

What it does

A preset strip across the top, a prompt box for anything the presets do not cover (usually with Randomize and Enhance), an apply/generate button that spends credits, and Accept & Close to commit the result back to the canvas.

What to expect

Presets replace each other rather than stacking. Accept & Close writes the result onto the canvas; closing without accepting leaves the canvas untouched but the generation still happened.

Use it well

Load a preset, read the prompt it wrote, edit that sentence rather than adding a second one.

What it does

You can read every control, load presets and understand exactly what you would get on any plan.

What to expect

The upgrade prompt only appears when you press the button that would spend credits.

Use it well

Open the locked labs once and read them. It is the cheapest way to know what a tier actually buys.

What it does

Upscale recovers resolution with an optional reference and custom prompt. UpMax is the heavy upscaler — preset-driven, with an extra context field and a creative reference mode.

What to expect

Upscale stays close to the source. UpMax will reinterpret detail it cannot recover, which is a feature when the source is soft and a hazard when it is not.

Use it well

Upscale for approved work, UpMax when the frame needs both size and rebuilt micro-detail.

What it does

Edit Lab is for structural, generative change — remove, replace, add, recolour or rebuild a region. Detect (left rail or Properties) runs on device and finds Subject, Background, Sky, Person and Object — click it on the main image — after a one-time ~4 MB model load, with no credits. Detect › Precise object selects exactly what you click, even things with no class name (a wheel, a badge, one car of several): load the precise model once under Local models, click to add points, use − (or Alt/right-click) to exclude, pick Mask 1–3, then Accept. It runs on this device and images are not uploaded. Range (left rail) builds selections from a luminance range or sampled colours (+ / − samples, tolerance, softness). Each selection is a stack you can Add, Subtract or Intersect, reorder, disable or solo — for example Object ∩ Colour range (red), minus a person.

What to expect

Selections are kept under Saved selections: rename, duplicate, hide, lock, invert, reuse or delete them across many edits. The prompt-mediated detector remains only as a clearly labelled, billed Fallback. Edit plan turns a sentence into a reviewable target, protections, action, preservation and candidate count; the normal preflight still picks the engine and shows the cost before anything runs. Photo-finishing requests (warmer, more contrast, vignette) show “This is a Develop adjustment” with Open in Develop or Continue anyway. With a selection active, Preservation review measures how much changed inside versus outside it, with a Difference view.

Use it well

Send to Develop moves the image and selection into a Develop mask; the sparkle button on a Develop mask sends the edited Develop image and the exact selection back. Each result shows how much changed inside and outside the selection, with Difference and Outside changes views. Develop is for appearance, Edit Lab for changing what is in the picture.

What it does

Create a palette, drop reference images into it, then press Train & extract look. The studio produces a style prompt, a negative prompt and a tag set.

What to expect

Signature strength climbs from Untrained through Training and Emerging to Signature as you accept or reject results. Use prompt applies the look; the palette can also be attached as a style chip in Playground.

Use it well

Four to eight references that agree with each other beat twenty that argue.

What it does

A browsable catalogue of preset looks with categories and search, separate from your trained palettes.

What to expect

Pinned looks sit in the sidebar style panel. Multi-select loads several at once.

Use it well

Pin the three you actually use; browse the rest only when a job needs a departure.

What it does

Create a look from a set of references, blend two looks into a third, and rename, tag and file looks alongside your other assets.

What to expect

A blend is a new look, not a live link — editing a parent does not update the blend.

Use it well

Blend a signature look with a neutral one when the signature is overpowering the subject.

What it does

A message thread anchored to the current image. Ask what is wrong with a frame, how to phrase a change, or what instrument to reach for.

What to expect

Replies are text. Nothing is applied to the canvas until you act on the advice yourself.

Use it well

Ask for the change as an instruction you can paste, not as an opinion.

What it does

The same panel switched into brief-led direction: state the intent and it routes you to the labs and instruments that serve it.

What to expect

The mode you last used is remembered locally between sessions.

Use it well

Use Atelier mode at the start of a job and Chat mode in the middle of one.

What it does

Input (Image, Prompt, View, Style Preset), Image (Generate Image, Edit Image, Enhance Image, Remove BG), Labs (Material, Form, Scene, Camera, Render, Style, Poster, Sketch), Video (Generate Video, Extend Video, Stitch Videos, Get Video Frame), Instruments (Camera Rig, Light Plan, Compose Shot, Motion Rig, Proportion Rig, Camera Preset, Rig Merge) and Output (Preview, Note).

What to expect

An Image node wired into a Generate node is used as the reference; upstream nodes execute before the node you pressed play on.

Use it well

Terminate every branch in a Preview node so you can see where a chain went wrong.

What it does

The camera, light, compose, motion and proportion gizmos rendered directly on the node card, emitting a rig on their output port.

What to expect

Rigs merge through Rig Merge and compile into the prompt clause at the generator. The gizmos are interactive on mobile through the step editor.

Use it well

Build one rig branch and fan it into several generators to keep a series consistent.

What it does

Name and save a graph, then reload it later. The workflow drawer lives inside the Workbench — a vertical tab on desktop, a bottom sheet on mobile.

What to expect

The graph persists across a reload of the session, so a refresh does not reset your wiring.

Use it well

Save a workflow the moment it produces one good result, not after the tenth tweak.

What it does

A browsing surface with two tabs: Library holds canonical templates that cannot be edited in place, Mine holds the graphs you saved.

What to expect

Opening a template loads a copy onto the canvas, so the original stays intact and your edits never overwrite it.

Use it well

Start from the nearest template and save it under your own name after the first good run.

What it does

Eight formats — Product Film, Brand Spot, Design Story, Concept Trailer, Social Cut, Music Piece, Walkthrough and Free Form — set the shot grammar, aspect and suggested references. You choose the runtime: a preset length or a custom 10–600 seconds, and the planner fits scenes and shots to it.

What to expect

A format is a starting point, not a lock; every field stays editable per shot. References are attached by role to the story (character, object, style, location) rather than to fixed shot numbers.

Use it well

Pick the format whose shot grammar you want, even if the subject differs.

What it does

Model (auto or a pinned engine, per project or per shot — Veo plus other video providers, and Higgsfield Cinema Studio where enabled and priced), Sound, Continuity from the previous shot's last frame, and a reference dock feeding look and subject into every shot.

What to expect

Ask for audio and the studio upgrades to a sound-capable model. Continuity hands the previous last frame in as the next first frame.

Use it well

Lock continuity for a walkthrough, break it for a montage.

What it does

A motion path with a mini-map and a speed curve describing how the camera moves through the shot — push, orbit, crane, track — and how it accelerates.

What to expect

The mini-map is a read-only echo drawn over the video stage so the planned move stays legible.

Use it well

One move per shot. Two moves in eight seconds reads as a mistake.

What it does

Renders run server-side. On reload, the studio reconciles anything outstanding: finished jobs are collected, and jobs that terminated without a usable result refund their credits.

What to expect

Jobs abandoned by a closed tab are swept after roughly twelve minutes. Completed Cinema renders auto-save to the gallery.

Use it well

If a clip seems stuck, reload rather than re-running — a second run is a second charge.

What it does

Render all queues the remaining shots after a confirmation that shows the expected cost. Stop queue prevents further submissions; a single shot or new take runs independently of the queue.

What to expect

Stop finishes the shot already rendering — provider jobs cannot be cancelled mid-flight — and then shows how many shots remain so you can resume. Leaving Cinema stops the queue the same way.

Use it well

Render one shot first to check the look, then queue the rest.

What it does

The dashboard lists your productions; Back returns to it from the production canvas. Projects are saved to your account with a local copy, and clips move to durable storage.

What to expect

Reopening Cinema recovers jobs that finished while you were away. A production with a shot rendering cannot be deleted until it finishes; rendered clips stay in your Library.

Use it well

Name productions by deliverable, not by date — the dashboard is your edit list.

What it does

A prompt composer with Draft, Standard and HD modes, raw and tile switches, and the usual stylize, chaos, weird, aspect and seed parameters. Prompts are capped at 2,048 characters.

What to expect

Each Create posts one job that returns a four-frame grid in the feed. The model is fixed at V8.1 — there is no version picker, and the V8 Edit Model and character reference are not available through our provider.

Use it well

Explore in Draft, then rerun the winning prompt in Standard or HD rather than upscaling a Draft frame.

What it does

Three reference roles: Image prompt (composition and content, up to 20), Style reference (look and palette, images only, up to 20) and Omni reference (carry one character or object across, max 1). Images can come from your computer or the Library.

What to expect

Each role is sent to Midjourney as its own parameter, so a style image never becomes the subject by accident.

Use it well

Use Omni for the object you need to keep and Style references for the finish — do not stack the same image in both.

What it does

A visual browser over 1,855 deduplicated SREF codes, 247 of them with real preview images from Lummi, SrefHunt and Kajimelo. A category rail on desktop (a strip on phone and tablet) and views for All, Favourites, Recents, My codes and With preview.

What to expect

Clicking a card body replaces the current style codes; the + button adds without replacing. On phone and tablet the first tap opens a detail sheet instead of applying. Codes without a preview show a labelled placeholder, and source pages open in a new tab.

Use it well

Filter to With preview when you need to see a look before spending; use Add to blend two codes, Replace to swap direction.

What it does

Apply profile or moodboard codes you already own with the Personalization picker, alongside SREF codes.

What to expect

Witness can only apply a code. Profiles and moodboards are trained on midjourney.com, not here.

Use it well

Save your own codes once so they appear under My codes and in recents.

What it does

Subtle and strong variation, Remix, Upscale and creative upscale, Remove background, Pan in four directions, Zoom out 1.5× and 2×, and Vary Region with a region picker.

What to expect

Every action is a new metered job that lands in the feed with a status row; use ← and → to move between frames in the viewer.

Use it well

Pick the frame first, then choose the action — actions always apply to the frame open in the viewer.

What it does

Finished frames are saved to the Master Gallery under the Lab source with the Midjourney model and prompt attached.

What to expect

The feed and the Library stay in step: deleting a job removes its frames from the Library, and deleting a frame from the Library removes it from the job.

Use it well

Favourite and tag in the Library — the feed is a working log, the Library is the record.

What it does

Each board holds image, note, file and frame cards. The shelf shows every board as a grid with search and sort by last edited, created, title or card count; duplicate, rename and delete live there.

What to expect

Boards autosave on every change to this device's browser storage. Adding new research results places them in free space — existing cards are never rearranged.

Use it well

One board per question ('rear lamp language', 'matte greens') beats one board per project.

What it does

Search your other boards and your Library, browse your own Pinterest boards and Pins once connected, and run web research where the web connector is set up.

What to expect

If a source is not configured the drawer says so instead of failing silently. Pinterest shows only your own boards and Pins. Saved cards keep their source link.

Use it well

Pull from your Library first — what you already made is the cheapest reference you have.

What it does

Select cards to group them into a frame, dedupe, shortlist, extract a palette or look, save a reference set with roles (subject, style, material, palette, composition, lighting) or drop them into the Dataset inbox.

What to expect

Frames group without moving the cards' content. Reference sets and the Dataset inbox are reachable from the tool rail later.

Use it well

Tag roles when you save a reference set so a later run knows which image is style and which is subject.

What it does

A chat dock that plans with you: it reads your request, can ask one clarifying question with chip answers, and proposes next steps for the board.

What to expect

It plans in text. Any image it suggests still renders through the normal metered engine path.

Use it well

Ask it to propose three directions from a selection, then render only the one you like.

What it does

Render or edit image cards in place with the image engines, including Higgsfield Marketing Studio where it is enabled for your account.

What to expect

Results are stored as image data, not temporary links, so cards survive reloads. Every board render is also saved to the Library under the Board source.

Use it well

Send to Playground when you need sweeps; stay on the board for one-off variations next to their references.

What it does

Choose Develop in the studio toolbar, press Develop on a canvas result, or press Open in Develop in the Library viewer.

What to expect

The image opens with the tool rail on the left, adjustments on the right and Images, Looks, Versions and History along the bottom. Your last edit is restored when you come back.

Use it well

Generate in Playground or MJ Studio, then finish the pick in Develop.

What it does

Light, Color, HSL, Curves, Grading and Detail panels each have numeric sliders; double-click a slider to reset it. Auto Light and the white-balance picker work from the image itself. “Edit with instructions” turns phrases such as “recover highlights but keep the background unchanged” or “make the paint richer and slightly cooler” into a numbered plan: each step lists its target, slider changes and a strength you can edit, and you apply steps one by one or all at once.

What to expect

The preview and histogram update as you drag. Every change is listed in History with Undo and Redo; an applied instruction plan undoes in one step. Steps that need a subject, sky, person, background or object mask show READY, NEEDS TARGET (click the photo), RUNNING or MODEL NOT LOADED; “Apply supported” applies the whole-image steps on their own. Gradients are only used if you tick “Allow approximate masks”.

Use it well

Hold the backslash key, or use Before/After, to compare against the original.

What it does

Crop offers Free, Original, 1:1, 4:3, 3:2, 16:9 and 9:16 with rotate, flip and straighten, plus Perspective & lens: vertical and horizontal perspective, distortion, aspect, scale and X/Y offset. Guided perspective (beta) lets you drag two guides onto lines that should be straight. Masks add brush, linear, radial, luminance-range and colour-range areas; colour ranges are sampled with the eyedropper and can include or exclude several colours. Each mask can add, subtract or intersect further components, and has its own exposure, contrast, highlights, shadows, temperature, tint, saturation and clarity. Show the result as a coloured overlay or black-and-white matte. Retouch heals a spot from a source area you can drag.

What to expect

Masks › Local Segmentation: press Install model once (about 4 MB). It then finds Sky, Person, Subject, Background and Object (click to select) on this device — your image is not sent anywhere for this. If no subject is clear, it asks you to click it. Detection follows common scene categories, so edges are approximate; use Feather, Invert and Add/Subtract/Intersect to refine. Face tools are not available yet.

Use it well

Use a luminance mask on the brightest tones and intersect it with a linear gradient to darken only a bright sky. To change what is in the picture rather than how it looks, press the sparkle button on a mask to edit that selection in Edit Lab; typing a request like “remove the person” offers Send to Edit Lab.

What it does

Overlays (image, text, shape, gradient) can be clicked on the image, dragged, resized from the corners and rotated from the top handle; arrow keys nudge and Delete removes. Looks apply a saved finish with a strength slider. Save your own, update or delete them, and share them as .wlook files. Snapshots under Versions store an edit you can return to. Export renders JPEG, PNG or WebP at the original size, 4K, 2K or a custom long edge.

What to expect

Save version adds a new Library item that records the edit and the source image. Send passes the finished image to Playground, Workbench or Board.

Use it well

Snapshot before trying a different Look so you can compare.

What it does

Marketing Studio generates and edits from a prompt and up to 16 references. Product Shots, Graphic Ads and Marketplace Design pair a product image with a live preset. Motion Transfer copies motion from a 4–30 s video onto your images. Cinema Studio renders directed shots with optional camera, light, era and palette controls.

What to expect

A rail switches between Apps, Workflows, Models and My Runs. Runs are saved on the server, so they keep going if you leave and come back.

Use it well

Start from a recipe app (Product Shots) before the raw provider model — the preset does the art direction.

What it does

Each run shows a live estimate in credits, then status, retry and cancel controls under My Runs.

What to expect

Finished images are saved to the Library under the Lab source with a 'higgsfield' tag, as real image data rather than an expiring link.

Use it well

Read the quote on every submit — it is computed from the actual request, not a flat rate.

What it does

Generate is a prompt panel with seed, steps, guidance and sampler. Canvas is a layered document with a staging area: results land as candidates over the canvas and nothing changes until you accept one.

What to expect

Accepted results are saved to the Library. Hosted Witness engines are available here too and are metered like anywhere else.

Use it well

Use staging to compare candidates in place before committing a layer.

What it does

One list over hosted engines, cloud Stable Diffusion engines and the models on your machine, with search across Hugging Face and Civitai for models to download to the agent.

What to expect

Cloud engines and LoRA training show a lock and the reason while that path is switched off. Downloads go to your machine only.

Use it well

Favourite the two or three models you actually use so the shelf stays short.

What it does

Shows what the local agent sees: device, models, install queue, machine settings and its activity log. Requests go straight to the agent on localhost.

What to expect

Nothing is uploaded and no credits are charged for a local render. If the agent is not running, the tab says disconnected rather than pretending.

Use it well

Start the agent before opening WebUI, then press refresh once to probe it.

What it does

Filter by source — canvas, playground, workbench, board, lab, cinema, story (older items) or upload — and search across prompts and tags. Favourite, star-rate and tag any item. Midjourney, Studio Apps and WebUI results are filed under Lab with their model and prompt.

What to expect

Bulk select gives download (each / zip / contact sheet), favourite, rate, tag, copy prompts and delete. Bulk upload downscales oversized images automatically.

Use it well

Tag at the moment of generation. Retro-tagging a thousand frames never happens.

What it does

Everything you own — palettes, styles, lab presets, workflows, looks and prompts — organised into category chips with rename, copy, use and delete.

What to expect

Assets are separate from media; deleting a render does not delete the look trained from it.

Use it well

Rename anything you will reuse. `Palette 4` is not a look name.

What it does

Installable packs priced in credits, installing into the surface they belong to — the style library, your palettes, a specific lab's presets, the prompt library or Playground.

What to expect

A pack can be removed and re-installed later without buying it again.

Use it well

Install one pack at a time so you can tell what changed in your preset strips.

What it does

Three modes: metadata only (device), Witness Cloud (2 GB bundled with any paid plan), or your own Google Drive routed into folders by type.

What to expect

A background daemon uploads and reconciles with a single leader tab, content is hashed so re-uploading is a no-op, and deleting asks whether to remove everywhere or only on this device.

Use it well

The library UI names your actual destination — if it says Drive, media is going to Drive, not to Witness Cloud.

What it does

Re-runs the current setup across a chosen number of images, varying the seed only.

What to expect

A queue with per-task status — Done, Failed or Cancelled — and progress on the running task.

Use it well

Use it once the prompt is right and only luck is left.

What it does

Generates a chosen number of variations around the current direction rather than repeating it.

What to expect

Broader spread than seed iteration; less comparable between results.

Use it well

Explore first, iterate second. Doing it in the other order wastes the good seed.

What it does

Paste prompts one per line, attach styles, and the panel shows Total Generations before the Run Batch button.

What to expect

The credit hint updates with the total. Nothing runs while another task is in flight.

Use it well

Prove the pairing with a 2 × 2 before you launch a 3 × 4.

What it does

Brief (state the intent in plain language), Autocraft (proportion and package), Lumen (light plan), Sigma (composition and shot), Seal (commit).

What to expect

Seal applies to canvas, exports the brief, or saves to the gallery. Use brief hands the compiled direction back to Canvas or Playground.

Use it well

Finish the brief before touching a gizmo. The stages are only useful in order.

What it does

Request a visual proof at any stage — a paid render showing where the brief currently lands. Projects are named, saved, duplicated and deleted inside the suite.

What to expect

A proof is a generation like any other and appears in the gallery.

Use it well

One proof after Autocraft, one after Sigma. Proofing every slider is expensive.

What it does

Format PNG or JPEG with a quality slider for JPEG, Resolution 1x / 2x / 4x, an Add Watermark option, and export every image in a pack as a zip.

What to expect

2x and 4x resample on the way out; they do not recover detail the render never had.

Use it well

PNG for hand-off, JPEG at high quality for review decks, watermark on for anything leaving the studio unfinished.

What it does

From the gallery selection bar: download each file, download a zip, or generate a contact sheet.

What to expect

On iPhone, single downloads go to Photos; zip archives are saved to Files.

Use it well

Contact sheets are the fastest way to review a sweep with someone who is not in the app.

What it does

A moderated community feed where members publish images and video.

What to expect

Posts go through automatic content moderation before they appear.

Use it well

Publishing is separate from exporting — a Wall post is not a deliverable.

What it does

Cost is computed server-side from the real request. Plan credits reset each period; top-up pack credits roll over for 60 days.

What to expect

A run that fails without a result refunds automatically. Hard caps apply per request: at most 4 images and at most 8 seconds of video.

Use it well

Set a spending alert in the account panel rather than watching the balance chip.

What it does

Studio-tier operators can supply their own provider API keys, stored encrypted and used in place of Witness metering for the engines they cover.

What to expect

Runs on your key are billed by that provider, not by Witness credits.

Use it well

Add a key only for the engine you run at volume; leave the rest on Witness metering.

What it does

Studio requires an account. While the early-access cohort is running, general sign-ups may be closed and you request access instead.

What to expect

Granted addresses are let in on their next sign-in attempt.

Use it well

Request access with the address you will actually work from — the grant is per address.

What it does

Turns a proportion drawing into a controlled hero frame without letting the engine reinvent the package.

What to expect

Geometry stays close to the drawing; finish and environment come from the prompt and the rig.

Use it well

Do not add a reference image on the first pass — you will not know which input moved the result.

What it does

Keeps an approved look while changing a named attribute.

What to expect

Composition off keeps the frame yours; the reference contributes surface and light.

Use it well

Write the prompt as a delta — 'same object, matte anodised finish' — not as a new brief.

What it does

Produces a coherent set rather than four unrelated images.

What to expect

Holding the style chip constant across seeds is what makes the set read as a set.

Use it well

Change composition between runs, never composition and style together.

What it does

Builds a comparable colour/material/finish matrix from one geometry.

What to expect

Because the seed and geometry are fixed, differences between frames are the finish.

Use it well

Tag as you go — the contact sheet is only useful if the labels survive.

What it does

Adds time to an approved frame without regenerating the subject.

What to expect

The still becomes the first frame; the move is described, not keyframed.

Use it well

Eight seconds is the cap — plan the move to land inside it.

What it does

Converts taste into a reusable, portable style source.

What to expect

The extracted negative prompt is doing as much work as the positive one.

Use it well

Reject aggressively early. A look trained on mixed feedback stays vague.

What it does

Uses the advisor to choose the operation, then performs the operation deliberately.

What to expect

Copilot returns text; nothing changes on the canvas until you run a lab yourself.

Use it well

Ask for one instruction, not a plan. Plans cost four generations.

What it does

Turns a proven manual sequence into a one-button pipeline.

What to expect

Upstream nodes execute before the node you pressed play on; typed ports refuse bad wires.

Use it well

Save the graph the first time it works, then iterate on a copy.

What it does

Searches a prompt × style space in one queued pass.

What to expect

The multiplication is charged in full — three prompts and two styles is six generations.

Use it well

Prove the pairing at 2 × 2 first.

What it does

Fixes the direction before spending on finished renders.

What to expect

Proofs are paid generations; the brief itself costs nothing.

Use it well

Two proofs per project. If you need more, the brief is not finished.

Objective

Preserve the architecture of a drawing while removing the drawing's own texture, line weight and paper.

What you are learning

You learn which sketch channel carries structure (Geometry, Contour) and which one smuggles in marker noise (Detail).

Try this

Load the sketch, turn Geometry and Contour on, Detail off, hold influence near 55, generate, then finish through Render Lab.

Objective

Use a reference purely as light and CMF intelligence instead of letting it dictate form or framing.

What you are learning

You learn that a reference is not one thing — it is three separable channels, and most of its value usually sits in two of them.

Try this

Attach the reference, enable Material and Lighting, disable Composition, and keep the subject entirely in the prompt.

Objective

Hold the design completely still and change only the photographic language around it.

What you are learning

You learn how much of a render's character is photography rather than design.

Try this

Take one approved frame into Camera Lab, change lens and height only, then re-run the same rig through Render Lab.

Objective

Progressively remove degrees of freedom, freezing form, then camera, then material, in that order.

What you are learning

You learn to stage a project so late decisions cannot undo early ones.

Try this

Only unlock the stage you are working on. Once form is signed off, never re-run the prompt that produced it.

Objective

Run the same reference three times, each time with a single channel enabled, to find out which attribute is actually valuable.

What you are learning

You learn to describe a reference in terms of what it contributes rather than what it depicts.

Try this

Keep prompt, seed and engine identical across all three runs and put the results side by side in the gallery.

Objective

Shoot one approved object from many camera positions and build a contact sheet without redesigning anything.

What you are learning

You learn that shot selection is a separate craft from design generation.

Try this

Work through the Compose presets in order and favourite as you go, then review the favourites as a set, not one by one.

Objective

Build a materials board from one frozen form by changing only finish and colour.

What you are learning

You learn how much perceived design change comes from CMF alone.

Try this

Fix the camera and lighting first. Change one material family per run and name each result in the prompt so the gallery search finds it.

Objective

Develop a design in small deliberate steps instead of rerolling for luck.

What you are learning

You learn the difference between evolution and random variation — and how far a locked seed will actually carry you.

Try this

One change per generation, written down. If a step makes it worse, step back rather than compounding.

Objective

Use productive failures as design stimuli instead of discarding them.

What you are learning

You learn to read outliers — the runs that misunderstood you are often the ones with an idea in them.

Try this

Hold everything constant except the seed, run Seed Iteration, then shortlist only the results you did not expect and push them through Remix.

Objective

Search a design space systematically instead of firing off random batches.

What you are learning

You learn to design the experiment before you spend the credits.

Try this

Write the axes first — for example 4 forms × 3 materials × 4 cameras = 48 observations — then run Batch Process and review by axis, not by image.

Objective

Turn a polished output back into exploratory DNA rather than treating it as an endpoint.

What you are learning

You learn that abstraction is a generative move, not a stylistic one.

Try this

Send the render through Sketch Lab, bring the line result back as a Canvas sketch input, then run Variations from there.

Objective

Reverse-engineer a visual language out of an image you admire and test it on something unrelated.

What you are learning

You learn to separate a look from the object that happened to be wearing it.

Try this

Extract the look, then apply it to a deliberately neutral subject — a sphere, a box, a plain product — before you use it on real work.

Objective

Convert something that is not a product — geology, folded paper, architecture, aircraft structure — into believable manufactured form.

What you are learning

You learn to transfer structural logic rather than surface appearance.

Try this

Abstract the source to line first. Never feed the photograph directly into the final render or you will get a collage.

Objective

Use a short move to discover static compositions you would not have framed by hand.

What you are learning

You learn that motion is a search tool for framing, not only an output format.

Try this

Keep the move slow and single-axis so most frames are usable, then scrub for the one frame you would have shot.

Objective

Extract the visual DNA shared by several unrelated references while avoiding copying any one subject.

What you are learning

You learn that a trained look built from four sources is far more transferable than one built from one.

Try this

Choose references that agree on light and finish but disagree on subject — the disagreement is what stops the copying.

Objective

Collide style sources from different disciplines to produce a second-generation synthetic design language.

What you are learning

You learn that new languages come from the seams between disciplines, not from within one.

Try this

Run each source separately in Playground first so you know its signature, then combine only the two that conflict most usefully.

Objective

Build believable design archaeology for a product that never existed.

What you are learning

You learn to hold three things consistent at once — product language, manufacturing era and photographic era.

Try this

Describe the period as a set of design and photographic attributes, for example '1970s industrial product language × rally-support mobility × documentary photography'.

Objective

Test whether a visual DNA survives across a product family or into an adjacent category.

What you are learning

You learn which cues are identity and which were just this object.

Try this

Fix the style palette, the camera treatment and the lighting; let only the product type change.

Objective

Select parents by hand and let the system produce descendants across several generations.

What you are learning

You learn that curation, not generation, is the scarce skill.

Try this

There is no single blend button — pair the selected outputs as references on one run, or run Remix on the stronger parent with the weaker one attached as a reference chip.

Objective

Let Auto-Explore run generations while you act only as curator between them.

What you are learning

You learn how quickly a search converges when a human selects between rounds.

Try this

Shortlist ruthlessly — three survivors per round maximum — and feed them back as references for the next round.

Objective

Reinterpret one concept across four future dates while protecting a named set of attributes.

What you are learning

You learn to separate true identity from present-day fashion.

Try this

Write the protected attributes down before the first run and refuse to change them, however tempting the result.

Objective

Test whether an object's identity survives four radically different contexts.

What you are learning

You learn which cues carry identity when the environment stops helping.

Try this

Change context per run and keep the camera treatment consistent so the object, not the photography, is what differs.

Objective

Use temporal motion as dynamic form analysis — a moving camera reveals surfaces a still never shows.

What you are learning

You learn to read a design in the round rather than from its best angle.

Try this

Push the trajectory further than you would ever present, then judge the frames where the form is least flattering.

Objective

Let motion propose compositions a static operator would not have chosen.

What you are learning

You learn that framing intuition has habits, and motion does not share them.

Try this

Use a slow drift rather than a dramatic move — the useful frames are usually mid-move.

Objective

Close the loop so an intermediate motion artefact becomes the input to the next design.

What you are learning

You learn that the workstation is circular; nothing is only an output.

Try this

Run the loop three times and keep every stage in the gallery so the lineage is legible afterwards.

Objective

Move intent out of prose and into instruments that can be adjusted one at a time.

What you are learning

You learn how little prompt text is needed once the instruments are carrying the intent.

Try this

Cut the prompt to the subject and one qualifier. Everything else goes into sketch, style, rig and reference.

Objective

Split one job into distinct roles: direction, design correction, photography, CMF, rendering and graphic treatment.

What you are learning

You learn that most bad results are one role doing another role's job.

Try this

Finish each role before starting the next, and name the role in your own notes so you know what you are judging.

Objective

Keep the product identity fixed while the campaign treatment changes around it.

What you are learning

You learn to build a campaign as layers on one asset instead of as new generations.

Try this

Approve and lock the product render first; every campaign variant starts from that same file.

Objective

Solve graphic treatment after the product design rather than mixing typography into the first generation.

What you are learning

You learn why typography that appears during generation is never the typography you wanted.

Try this

Generate with No Text, then add every word deliberately in Poster Lab.

Objective

Build a coherent editorial set from a single approved design, as a photographer would shoot a day.

What you are learning

You learn to plan coverage — hero, detail, context, alternate — instead of generating until something works.

Try this

Write the shot list before you generate, then work it in order and favourite as you go.

Why it happens

The selected engine has no image-reference support, or the chip influence is at the bottom of its range.

How to confirm

Open the model library — capability warnings state up front when a model ignores image references.

How to fix

Switch to a reference-capable engine, or raise the chip influence by dragging up on the thumbnail.

Why it happens

A style chip at high influence, or a reference with Composition enabled, is pulling shape as well as look.

How to confirm

Drop influence to roughly 30% and re-run on the same seed; if the subject returns, influence was the cause.

How to fix

Material and Lighting on, Composition off. Use a separate subject chip for identity.

Why it happens

A style preset is asking for text — poster and editorial looks frequently do.

How to confirm

Clear the style preset and re-run on the same seed. If the text disappears, the preset was the source.

How to fix

Switch the guided prompt builder to No Text, or use the Text / No Text flag in the Playground prompt bar.

Why it happens

Some engines cannot accept image references or sound; the studio selects a capable model rather than dropping your input.

How to confirm

The model library flags the capability, and the run reports the model actually used.

How to fix

Pin the model explicitly if the engine matters more than the reference.

Why it happens

Aspect ratio is a generation parameter, not a crop — the engine re-frames the subject for the new canvas.

How to confirm

Re-run the same seed at the original ratio; if the composition returns, the ratio was recomposing it.

How to fix

Decide the ratio before you lock a seed, or use the Compose crop lock instead.

Why it happens

The rig only affects the next generation; Camera Lab re-shoots the existing frame as a new paid run.

How to confirm

A rig with untouched instruments contributes no clause at all — the chip shows what it will send.

How to fix

Touch at least one instrument before expecting the rig to change anything.

Why it happens

A preset and a palette-trained look are two style sources applied to one run.

How to confirm

Remove one and re-run on the same seed to see which is dominating.

How to fix

One style source per run. Blend two looks into a third if you want both.

Why it happens

Generation resolution, Upscale and Export Scale are three different things: the first decides detail, the second rebuilds it, the third only resamples.

How to confirm

Exporting at 4x from a 1K render gives a large soft file.

How to fix

Generate at the highest class your plan allows, then Upscale or UpMax, and export at 1x.

Why it happens

Copilot messages are advisory, but every lab run or regeneration you perform on its advice is a paid generation.

How to confirm

The account usage breakdown attributes the charge to the generation, not the chat.

How to fix

Ask for one instruction and run it once.

Why it happens

Browse-then-gate: labs open on every plan and only the credit-spending action is tier-gated.

How to confirm

The upgrade prompt appears on the generate/apply button, not on open.

How to fix

Use the open lab to read exactly what the tier buys before upgrading.

Why it happens

Media holds generations; Assets holds palettes, styles, presets, workflows, looks and prompts.

How to confirm

Filter Media by source — canvas, playground, workbench, board, lab, cinema, story (older items) or upload — before assuming it is gone.

How to fix

Search across prompts and tags rather than scrolling.

Why it happens

Canvas, Playground and Workbench are separate surfaces, but they write to one shared library.

How to confirm

Anything generated anywhere appears in Media with its source tag.

How to fix

Use the source filter as your map back to whichever surface made the frame.

Why it happens

Video runs as a server-side job; a closed or reloaded tab does not cancel it.

How to confirm

On reload the studio reconciles outstanding jobs, collects finished renders, and refunds jobs that terminated without a usable result. Abandoned jobs are swept after roughly twelve minutes.

How to fix

Reload once and wait. Re-running is a second charge for the same shot.

Why it happens

Storage mode is metadata-only, so nothing but records left the original machine.

How to confirm

The library header names your actual destination — device, Witness Cloud or Drive.

How to fix

Turn on Witness Cloud or connect Drive, then run Sync device to merge the libraries.