MetaBrain Help & Guide

MetaBrain writes sale-ready metadata for microstock: an SEO-optimised title, a description, and a ranked keyword set for every photo, video, or vector you drop in.^* One file costs one credit. What used to take ten minutes of manual tagging per file takes seconds.

1. Open metabrain.online/studio in any modern browser (Mac, Windows, Linux), or launch the macOS desktop App.

2. Sign in with Google.

3. Drop your files. That is genuinely all: Mode A is selected by default and needs zero preparation — no reference images, no keyword blocks, no settings. In Studio, processing starts automatically the moment files land (two files at a time); in the App, press the main action button.

4. Review the result cards, edit anything you like, export.

Everything else on this page — Modes B/C/D, keyword blocks, trend scores — is optional tooling for power users. You can ship your first batch without reading further.

Both share the same AI backend and produce identical quality. Studio (web) — no installation, runs everywhere, and includes AI brand detection (Detect IP). App (macOS) — adds direct metadata embedding into your files (Write to Files), the IP Scanner badge, and native folder drag-and-drop. Keyword blocks, learned corrections, and history sync between the two automatically.

One processed file = one credit. Free — 30 credits at signup, no card required. Hobby — 300 credits/month ($9). Pro — 1,100 credits/month ($19). Credits Pack — +500 for $15, one-time, never expires. The first Ground-Truth correction of each image is free (see "Steering the AI").

Mode A is the default working mode and the one you will use most of the time. It needs nothing but your files — no reference images, no keyword blocks, no configuration. Drop a photo, a video, or a vector: the AI reads the actual content and writes a complete metadata set from scratch.^*

If you remember one thing from this guide: open Studio, stay in Mode A, drop files.

The sidebar accepts up to 10 reference images. This is an optional fine-tuning tool: references nudge the keyword vocabulary towards the tone of a series — documentary, commercial, fine-art — without changing what the AI sees in the main file.^* Most users never need them, and results without references are in no way inferior: the AI simply works from the image itself.

Single files, small mixed batches, and anything that does not fall into a clear repeating niche. For high-volume niche production, see Modes B and D. For a series of near-identical shots, see Mode C.

A ranked keyword set (up to 49 by default — adjustable from 20 to 80 in Output Settings), a sentence-case title, and a factual description built around buyer intent.^*

Mode B layers your own keyword library on top of the AI. A block is a themed word list — "Food Photography", "Travel — Mountains", "Business Technology". For each image, MetaBrain automatically picks the best-matching block^* and blends its vocabulary into generation. The payoff: consistent niche wording across your whole catalogue.

Studio has a built-in block editor in the sidebar: create a block, give it a descriptive English theme name, add keywords, save. Blocks live in your account and sync automatically with the desktop App — build your library once, use it everywhere.

The desktop App also imports plain .txt files — one keyword per line. The filename is the theme name, so use descriptive English noun phrases: "Business portrait team.txt", never "block1.txt". An optional first line sharpens matching between similar-looking niches:

Keywords are lowercased and de-duplicated on import.

A block is selected by its theme name, so that name must describe the theme precisely and specifically — "Aerial nature landscape", "Business handshake in office" — never vague or garbled names like "block1", "misc", or a random string of words. The AI matches every image against these names: a vague, generic, or misleading name makes the block attach to the wrong images or be skipped entirely (no match). In the App the theme name is the .txt filename; in Studio it is the block name you type. Getting the name right is the single biggest factor in whether a block helps or hurts your results.

Use single words per line — enter "business" and "meeting" as two separate lines, not "business meeting". The AI builds effective two-word phrases on its own where they help search performance.

If no block fits the image, MetaBrain falls back to pure AI analysis — the same quality you would get in Mode A.^* Nothing breaks; you simply get an unassisted result.

Photographers with established niches who process many similar images regularly. Build blocks once, reuse forever — Mode B is the most efficient path for high-volume single-image workflows.

Mode C is built for batches of near-identical shots — the same scene from slightly different angles, moments, or lighting. It rotates a unique Focus Point (mood, composition, subject, colour) into each image's prompt, so every title and description in the batch comes out distinct^* — which protects you from duplicate-content rejections at the agencies.

You have 3–50 shots from one session that are visually similar but all need to be submitted: bracketed exposures, model poses, sports sequences, food styling variations.

1. Load keyword blocks in the sidebar (same library as Mode B).

2. Drop the batch of similar images into the grid.

3. Start processing.

Each image gets its own perspective variation woven into the prompt — titles and descriptions differ even for nearly identical frames.^*

Unique titles and descriptions across the whole batch; keyword count follows your Output Settings (up to 49 by default).

Mode D processes a large batch of unrelated images at once. Each image gets its own fully independent title, description, and keyword set.^* Keyword blocks from the sidebar act as vocabulary pools the AI draws from when relevant.^*

The workhorse for production: process an entire shoot or a catalogue dump in one session. There is no upper limit on batch size.

Drag any thumbnail onto another to reorder the batch before or after processing — the grid updates instantly. Use it to group similar shots or restore your original shoot sequence.

The A→Z button sorts all result cards alphabetically by filename — handy for matching results to your folder structure.

Double-click any card to open the detail view: full-size preview, Category selector, and Title/Description fields. Generate Title & Description runs AI on that single image. Edits made in the popup are saved back to the card — and in Mode D, pressing Save in this window is what writes the card to your History.

The Regenerate button on any card re-runs AI for just that image without touching the rest of the batch.^*

With Editorial Mode enabled in Settings (available in both the App and Studio), dropping files shows an interactive pre-selection grid before AI processing starts. You mark which files should be treated as editorial — the rest are processed as standard commercial images.

1. Enable Editorial Mode in Settings.

2. Drop your files.

3. A grid appears — click any thumbnail to toggle its editorial status.

4. Use Select All / Clear for bulk marking.

5. Start the analysis — marked files get editorial metadata, unmarked files get standard commercial metadata.

You can drop additional files while the grid is open; current selections are preserved, and duplicates (same path + size) are skipped automatically.

Editorial-marked files receive a different AI treatment:

- The capture date (from EXIF) is added to the title

- The description is written as a factual who/what/where caption rather than commercial copy

- Keywords lean towards documentation and reportage

- The Editorial flag is set automatically in all CSV exports

Settings → Output controls the shape of every generation, in both Studio and the App:

- Title — cap by word count (5–15 words) or switch to a character-based cap for agencies with strict character limits

- Description — target length in characters

- Keywords — target count, from 20 to 80 (default 49)

By default the keyword number is a ceiling, not a quota: MetaBrain will not pad the list with weak filler just to hit a number.^* If your workflow requires an exact count, enable the strict-count option and the AI will deliver exactly the number you set.

Each generated keyword carries a relevance score from 0 to 100 — how central it is to what a buyer would type when searching for this exact file.^* The chip shading follows the score: bright chips (80+) are core terms, mid-tone chips (50–79) are supporting long-tail terms, and chips below your threshold are dimmed.

The threshold slider in the sidebar sets your personal cut-off. Chips below it dim, and the [ N ] counter on each card shows how many keywords pass. Use it to triage the weak tail at a glance: review the dim chips, keep the useful ones, delete the rest before export.

The AI makes a first guess about what is in an image — and sometimes it guesses wrong. You can override it. On any finished result card, click Regenerate (App) / Refresh (Studio) to open the Ground Truth panel: a free-text box where you state exactly what the AI misread. Your words become the source of truth for the next generation.

Type any fact the AI got wrong — for example:^*

- Demographics — age, gender, ethnicity, number of people ("two senior women, not one young man")

- Emotion — "anxious, not happy"

- Setting and location — "hospital waiting room in Lisbon, not a generic office"

- Commercial intent — what the shot is really *for* ("about online banking security, not just a person at a laptop")

These facts OVERRIDE the visual interpretation. The AI rewrites the title and description and re-ranks the keywords as if your facts are true — with no hedging ("appears to", "seems") and without contradicting you.

A correction changes only what you flagged. Everything the previous result got right is kept — the full keyword count, correctly-read emotion and composition, lighting and setting. You will not lose a good result by fixing one detail. Leave the box empty and click Regenerate to simply re-roll the same image.

In the batch modes, tick Apply to whole theme before regenerating. The same correction is pushed to every image that shares that keyword block, and they are all regenerated together — fix one misread, fix the whole set.

The first Ground Truth regeneration of each image does not cost a credit. While it is available, a green First correction free badge sits next to the Regenerate button. After you spend it, further regenerations of that same image cost one credit, as usual.

Correction Intelligence (CI) is a learning system that watches how you edit AI results and turns your edits into personal rules that improve future generations. The more you correct, the better MetaBrain fits your niche and taste.

Edit a result — delete keywords you never want, add ones the AI missed — then press Learn on the card (available in both the App and Studio). Your changes are saved as correction rules and injected into future AI requests automatically. In the App, `Cmd+Enter` learns from all visible cards at once.

Global rules — apply to every image. Example: you always delete "candid" — CI stops generating it.

Theme rules — apply when the image matches a keyword-block theme. Words you consistently remove from a theme get banned; words you add get suggested.

Context rules — used where there are no blocks (Mode A): MetaBrain derives the context from the top keywords and groups corrections across similar contexts.

Settings → Correction Stats shows your full rule dashboard: Frequent Hallucinations (keywords the AI keeps generating that you keep deleting — blacklisted), Missing Words (keywords you add that the AI misses), and an AI Accuracy Trend sparkline. Click any entry to review or delete it; you can also reset individual rules or everything. In Studio, the Corrections page serves the same purpose.

Stock agencies rank results by relevance to a buyer's search query. Buyers search for what they need to communicate, not what they see in a photo.

❌ Wrong approach: "woman, smartphone, kitchen, smiling"

✅ Right approach: "adult woman making online payment at home kitchen"

MetaBrain is calibrated to think about buyer intent first. Every title answers WHO does WHAT, WHY, and WHERE. Every keyword set prioritises commercial use cases over visual inventory.

The filename is read as a trusted hint — especially for places. A file named "lisbon-alfama-sunset.jpg" puts Lisbon into the title and keywords; "IMG_4382.jpg" tells the AI nothing. Locations are exactly what the AI cannot guess from pixels alone, so name files with the real place, event, or subject whenever you can.

Stock search engines index individual words as well as phrases. MetaBrain outputs keyword phrases to maximise surface area:

`"online banking"` → indexed as "online" AND "banking" AND "online banking"

This multiplies the number of searches your file can appear in. Do not break MetaBrain's phrases apart — the comma-separated format is intentional.

MetaBrain assigns one of 18 standard stock categories (e.g. "Business & Finance", "Landscape & Nature") automatically; you can override it on the card. Correct category assignment noticeably improves placement in agency browse pages.

Toggle Editorial for images that cannot be licensed commercially — recognisable people without model releases, branded products, news events. This prevents rejections for unlicensed commercial use.

Click the × on any chip to remove it. Click the + Add keyword area at the end of the list, type, and confirm with Enter or a comma. Double-click a chip to edit its text in place.

Many keywords come out as two-word phrases ("family cooking", "healthy eating") rather than single words. This is intentional: buyers search with phrases, and a phrase matches their query precisely, while a bare noun like "family" competes with millions of files. Concept phrases occupy the top positions, where agency algorithms weigh relevance most; broad single words sit lower to cover short queries. Adobe Stock, Shutterstock, and Getty fully support multi-word keywords. Tags longer than two words can look like spam to reviewers, so MetaBrain caps every keyword at two words.

Drag any chip to a new position — position order is what agencies see. The first 10 positions carry the most search weight; the App highlights them as gold Power Keywords. If something needs promoting, drag it into the top ten.

Click directly on the title or description text to edit it inline. Press Enter or click elsewhere to confirm.

Every card includes a Category field. The AI picks one of 18 standard microstock categories automatically; click it to choose another. The category is included in all export formats: CSV, Word, XMP, and Clipboard.

In the App, each card has a checkbox. Select several cards to Apply Theme (inject a keyword block into all of them) or Regenerate them together (a confirmation dialog warns that manual edits will be lost). Select All / Deselect All live in the sidebar.

`Cmd+Z` — undo on the last modified card

`Cmd+Enter` — learn from all visible cards (App)

`Escape` — cancel the current edit

In the App, every result card has a round ↺ button. Each card stores up to 5 undo steps, independent of other cards. A snapshot is captured automatically before AI generation, before you edit the title or description, and before any chip change. In Mode D, the ↺ button lives in the detail popup.

In Studio, the ↺ Reset button on each card returns the card to the original AI output in one step — useful when an editing session went sideways and you want the clean starting point back.

`Cmd+Z` — undo on the last card you modified (Modes A/B/C) or on all selected thumbnails (Mode D, App). Does not interfere with native `Cmd+Z` inside text fields.

Before AI processing begins, MetaBrain runs a lightweight technical analysis on each image, in a background thread that never blocks the UI. Results appear as coloured badges on result cards.

Resolution — image should be ≥ 4 MP. 4–8 MP triggers a warning; below 4 MP is an error. Shutterstock requires 4 MP minimum; Alamy and Getty prefer 8 MP+.

Sharpness — measures edge contrast in the centre crop. Low score = blurry subject (camera shake, missed focus, heavy noise reduction). Artistic bokeh may also trigger it — use your judgement.

Exposure — counts blown highlights and clipped shadows; both are hard to fix in post and a common rejection reason.

Noise — measures digital grain, typically from high ISO. Excessive noise is a frequent rejection trigger at larger print sizes.

Colour Cast — detects a dominant tint from incorrect white balance. Strong casts — especially in skin tones — are a rejection risk.

Error (✗) — value is outside what agencies typically accept. High rejection probability; fix or exclude before uploading.

Warning (⚠) — borderline value. The image may still pass, but review it manually. Conservative agencies (Getty, Alamy) apply stricter standards than volume platforms.

Modes A/B/C — a red ✗ N issues or amber ⚠ N warnings badge on the card. Mode D — a compact badge on the thumbnail; double-click to see the full table with values.

The pre-scan is advisory only — MetaBrain never blocks or skips files that fail checks.

An optional tool in the App (Settings → Tools → IP Scanner). When enabled, the AI checks each image for identifiable brands, trademarks, licence plates, and named establishments. If any are found, an amber ⚠ Editorial badge appears on the card — hover for a summary, click to see the detected objects overlaid on the thumbnail. With the scanner off, the badge is simply hidden.

In Studio, every finished card has a Detect IP button: a free scan for logos, brand marks, licence plates, and named establishments that would require a property release (JPEG, PNG, WebP). Found objects are drawn as orange boxes directly on the thumbnail; View list shows every detected item by name, and you can re-run the scan after editing the image. If nothing is found, you get an explicit "No IP issues found".

Use the result to decide: submit as editorial, retouch the marked areas in your editor, or exclude the file.

Editorial content can be sold for news, educational, or documentary use, but not for advertising. Most agencies require a property release or an editorial licence for images with identifiable brands, storefronts, or public figures. All IP tools are informational — MetaBrain never rejects or modifies your files without you asking; you decide how to licence and submit.

Every result card has an Inspector button that audits the metadata against stock-agency requirements: title and description length, risky or banned terms, duplicate keywords — and spelling. A 46,000-word dictionary catches garbled keywords, including subtle cases like a missing or an extra trailing letter that spell checkers usually miss.

The button shows an issue counter; ✓ All clear means nothing was found. Fix flagged issues before export — misspelled keywords are dead weight (nobody searches for them), and rule violations are avoidable rejections.

Trend Score is an optional tool (Settings → Tools → Trend Score). When enabled, a commercial-potential indicator appears on every result card. The score models earning potential = demand × (1 − market saturation): oversupplied commodity subjects (apple on white, generic handshake) score low even when the topic is popular.

① Gradient bar — the filled portion shows the score out of 100.

② Numeric score — colour-coded from bright (high) to muted (low).

③ Momentum arrow — ↑ Rising · → Stable · ↓ Declining

80–100 — Hot — high demand, trending now

60–79 — Rising — growing commercial interest

40–59 — Stable — consistent evergreen demand

0–39 — Niche / Declining — limited or shrinking audience

Click the score bar to open the full Trend Panel for that image:

- Agency Brief matches — how many current agency shoot briefs this image's keywords match

- Demand keywords — high-demand terms already in your set

- Avoid zone — oversaturated keywords to remove or deprioritise

- Category saturation — how crowded this content category is

- Brief list — matched briefs with their subjects and freshness date

- Recommendations (+1 credit) — a personalised list of trending keywords to add to this specific image

Availability: the Trend Score itself is included on every plan, Free included. The brief with recommendations (the Recommendations button and the PDF brief) is available on Hobby and above.

Hover over the bar (without clicking) for a compact tooltip with 2–3 high-demand keywords not yet in your list.

Trend Score is an estimate — it maps content category and keyword signals to microstock demand patterns, not real-time sales data. Verify against your own agency dashboards before making production decisions.

MP4 and MOV work everywhere. The desktop App additionally handles AVI, MKV, and WebM. In Studio you can also drag in WMV, M4V, MTS, and MXF — as long as your browser can decode the codec inside.

MetaBrain extracts 5 frames across the timeline (at 10%, 30%, 50%, 70%, and 90%) and assembles them into a single collage. The AI analyses the collage to understand the full arc of the video^* — not just one frozen frame — and writes footage-appropriate titles and keywords. One video costs one credit, same as an image.

No browser can decode ProRes, DNxHD or CineForm, and that is a limit of the browser itself, not of Studio — the error message names the exact codec it found. The quickest way round it is the H.264 proxy your editing suite already writes: drop the proxy, then set Master file name in the export panel to the master's extension (mov, mxf) so the CSV names the file you actually upload. Otherwise transcode a copy with Shutter Encoder or HandBrake (both free, Windows/macOS/Linux), or process the master in the desktop App.

Videos take longer than images — the frames must be extracted and a larger payload analysed, so expect up to a few minutes for long clips; keep the tab open. Heavily edited clips with rapid cuts may produce lower-confidence results — consider processing a representative segment separately.

Everything on this page works today. What is still growing: key detection needs a human on roughly one track in three (the field is always editable, and a key you pick by hand is the one that reaches the CSV), the shelf table behind the music Trend covers 104 measured phrases and keeps growing, and the catalogue analysis is App-only for now. Nothing is hidden behind the badge — it is there so you check the key and the loudness before you submit rather than after.

MetaBrain treats audio as music, not as a picture with a soundtrack. WAV, AIFF, MP3, M4A, FLAC and OGG are accepted. Before the AI sees anything, the track is MEASURED on your own machine: length and bar count, tempo, key and its Camelot code, integrated loudness (LUFS), loudness range, true peak (dBTP), sample rate, bit depth and channels. Those measurements are handed to the model as facts, so it describes your music instead of guessing numbers about it. What comes back is a title, a description and keywords written the way music is actually bought — by use case, mood, instrument and tempo. One track costs one credit, same as an image. Audio is processed in Mode A — in the desktop App and in Studio alike; the other modes work from keyword blocks and batch grids built for pictures.

Under the description you get the strip: LEN with bar count, BPM with a one-click swap to the half/double reading, KEY as an editable picker with the Camelot code beside it, LUFS, LRA, PEAK and FMT. Everything measured is editable, because you know your own track better than any estimator does. Coloured chips name what a library would reject the file for — clipping, true peak over the limit, mono, DC offset, loudness far from the shelf — and each chip opens into four lines: what is wrong, where in the track, the fix with the number computed for your file, and which platforms actually care.

Honestly measured rather than claimed: 7 of 8 correct (88%) against the MIDI of real projects, and 60.4% exact / 69.0 weighted on the public GiantSteps set of 604 expert-annotated tracks — ahead of KeyFinder, behind Mixed In Key. Roughly one track in three needs a human, and nothing in the confidence number predicts which one, so the field is simply always editable. A key you pick by hand is treated as certain and goes to the CSV; a low-confidence estimate is left out of the export rather than shipped as a wrong number.

Two folded tabs sit under the strip. Stock holds what a library asks for — category, genre, mood, instruments, vocals, use cases. Release holds what a distributor asks for — artist, ISRC, ISWC, writer splits, IPI, publisher, year, explicit flag and the AI disclosure that stores have shown in credits since April 2026. Anything your DAW already wrote into the file (artist, ISRC) is read from the file and never overwritten by the model. The cover slot checks the image against every store's rules at once: square, at least 1600×1600 for TuneCore, 3000×3000 recommended everywhere. Cuts bounce 15, 30 and 60-second versions aligned to the tempo grid, rounded down to whole bars so they end musically.

Write to Files puts the metadata inside the audio itself: ID3v2.4 for MP3/WAV/AIFF, Vorbis comments for FLAC/OGG, MP4 atoms for M4A, plus RIFF LIST-INFO (what Finder and Windows columns read) and a broadcast BWF chunk. An empty field never erases what you already had. For the agencies, the export menu holds one sheet per platform: Pond5 Music, Storyblocks Audio, Motion Array Audio, AudioJungle (Envato) and a distributor release sheet for DistroKid, TuneCore, CD Baby or Amuse. On a batch of tracks the picture agencies disappear from the menu, and on a batch of photos the music sheets do — you only ever see the sheets that fit what you loaded.

Pond5 takes WAV or AIFF only, stereo, under 10 minutes, and forbids spaces and dashes in the file name; it also needs a separate FTP password set in your dashboard. Storyblocks takes WAV or AIFF from 60 seconds — and their music licence is exclusive, so a track sent there cannot be sold anywhere else; the export menu marks that sheet so you know before the click. Motion Array wants a WAV 24-bit/48 kHz master with an MP3 preview and is non-exclusive. AudioJungle wants everything at 44.1 kHz/16-bit stereo and the preview as MP3 320 CBR. Envato paused new author sign-ups in 2021 and it is still paused — the sheet ships anyway for the authors already inside.

Envato is the only library that makes the author watermark the preview, and their rules are exact: required on anything two seconds or longer, heard every 10 seconds over 20 s, every 5 seconds on 10–20 s, every 2–3 seconds on 5–10 s, and once on anything shorter — with the sample unaltered, no pitch changes and no effects. AudioJungle watermark on previews… in the export menu does exactly that. We do not ship Envato's watermark file: it is their asset and they hand it to their own authors, so you point at your copy once and MetaBrain remembers it. Your master is never touched — the watermark goes on a separate preview, written as MP3 320 CBR, the format their upload form asks for. Convert tracks to MP3 320… does the same conversion on its own for any track.

A track has no subject, so the picture Trend Score means nothing for it. The music engine answers a different question — what to do with this track next — and every line comes from a measurement or the calendar: short versions a library licenses that you have not cut yet; loudness sitting below the −14…−9 LUFS band everything else is auditioned at, with the exact gap in dB; how many use-case words a brief could match on; your tempo against the band the genre is actually searched at; a seasonal window, because a Christmas cue uploaded in December missed its year and one uploaded in September did not; and how full the shelf your keywords name already is. Shelf sizes are measured on AudioJungle rather than guessed, and the advice only ever points at another word you chose — an empty shelf is not an opportunity if nobody searches for it.

Analyse a folder of tracks… in the export menu measures a whole folder of masters and maps the catalogue: which tempo and length cells are empty, whether everything sits in one key mode, how far levels jump between tracks, which seasons you have nothing for and when the nearest window opens, and how much of the catalogue is stacked on one crowded shelf. It reads the files, not our history, so it works on a catalogue you built long before you found MetaBrain. Files it could not measure are counted and named rather than quietly dropped. Expect about two and a half seconds per compressed track.

Instead of the cloud, the desktop App can send your files to an AI model running on your own Mac — no internet needed for processing. Your images never leave the computer, no credits are consumed, and generation keeps working when your connection drops.

Available in the macOS App only (not Studio), on the Pro plan. You need one helper program that runs the model — the easiest is LM Studio (everything by mouse, no Terminal), the alternative is Ollama.

First check how much memory your Mac has: Apple menu () → About This Mac → the Memory line. The rule of thumb: the model should take up no more than half of your Mac's memory. All models below are vision models (they understand images):

- 8 GB — Qwen2.5-VL 3B, the only comfortable option. LM Studio search: `qwen2.5-vl-3b` · Ollama: `ollama pull qwen2.5vl:3b`. Close other apps while processing.

- 16 GB — Qwen2.5-VL 7B, the recommended choice for most users. LM Studio: `qwen2.5-vl-7b` · Ollama: `ollama pull qwen2.5vl:7b`.

- 32 GB — Gemma 3 12B, noticeably better quality. LM Studio: `gemma-3-12b` · Ollama: `ollama pull gemma3:12b`. Or stay on 7B if you prefer speed.

- 64 GB and up — Qwen2.5-VL 32B or Gemma 3 27B, the best quality a local model can give, but slower. LM Studio: `qwen2.5-vl-32b` · Ollama: `ollama pull qwen2.5vl:32b`.

The newer the chip (M1 → M4) and the higher its class (Pro / Max), the faster the generation. When in doubt, start small — you can always download a bigger model later.

1. Go to lmstudio.ai, download LM Studio for Mac, drag it into Applications and open it.

2. Click the magnifying glass (Search) and type the name of the model picked for your Mac above — for most users that is `qwen2.5-vl-7b`. Download the MLX 4-bit build.

3. Open the Developer tab (the `>_` icon in the left sidebar). In the selector at the top, pick the model you just downloaded so it loads — this step is easy to miss, and without it the server is empty.

4. Press Start Server. You should see it running at `http://localhost:1234`.

5. In MetaBrain: Settings → Local AI model → toggle on. Leave Endpoint URL exactly as pre-filled (`http://localhost:1234/v1`) and leave Model name empty.

Check it works: drop one photo and process it. If your credit counter does not go down, the file was processed on your Mac.

1. Install Ollama from ollama.com and open it — its server starts automatically in the background.

2. Open the Terminal app and run the Ollama command for your Mac from the list above, e.g. `ollama pull qwen2.5vl:7b`. Wait for the download to finish.

3. In MetaBrain settings set Endpoint URL to `http://localhost:11434/v1` and Model name to exactly what you pulled, e.g. `qwen2.5vl:7b` — with Ollama this field is required.

Locally: Modes A, B, and C, Regenerate and Ground Truth corrections, vectors (converted on-device), and videos — the 5-frame collage is built on your Mac and analysed by your local model.

Still cloud-based: Trend Score (local results simply have no trend badge), Mode D batch-grid analysis, and the Correction Intelligence advisor. Signing in and syncing history or keyword blocks also needs a connection — but if the connection drops mid-session, local processing keeps going.

Expect 1–3 minutes per file on a laptop — the whole model runs on your processor; that is normal. A 7B model is also noticeably weaker than the cloud pipeline with its extra quality gates — expect shorter keyword sets and more manual review.^* The cloud remains the recommended default for final production work; the local model shines for drafts, sensitive material, and offline sessions. Before a session, close heavy apps — the model needs 5–6 GB of free memory.

- "Endpoint not reachable" — the local server is not running. LM Studio: Developer tab → Start Server. Ollama: open the app.

- "Did not return valid JSON metadata" — the model is text-only or too small. Use a vision model such as Qwen2.5-VL.

- Very slow or the Mac freezes? Take the smaller 3B model.

Copies the title, category, description, and all keywords for a single image as formatted text. Paste directly into the agency upload form.

Exports results as a CSV file formatted exactly for a specific stock agency. Available profiles:

- Shutterstock — Filename, Description, Keywords, Categories, Illustration, Editorial

- Adobe Stock — Filename, Title, Keywords, Category (numeric), Releases

- Getty Creative (ESP) — 13-column ESP: file name, title, description, keywords, location fields

- Getty Creative Video (ESP) — same ESP; poster timecode instead of created date, no color column

- Getty Editorial (ESP) — same ESP + personalities field (names of recognizable people); editorial caption

- iStock (ESP) — identical to Getty Creative Image

- iStock Creative Video (ESP) — identical to Getty Creative Video

- iStock Creative Illustration (ESP) — identical to Getty Creative Image

- iStock Editorial Image (ESP) — identical to Getty Editorial

- Alamy — Alamy Image Ref ⚠️, Filename, Caption, Tags (semicolon-separated), Date Taken, Location, Country

- Dreamstime — Filename, Image Name, Description, Category 1–3, keywords (quoted), Free, W-EL, P-EL, SR-EL, SR-Price, Editorial, MR doc Ids, Pr Docs — upload CSV only AFTER images appear in Unfinished

- Depositphotos — Filename, Description, Keywords, Editorial, Nudity

- Pond5 — OriginalFilename, Title, Description (max 500), Keywords, Price, Editorial, City, Region, Country

- 123RF — oldfilename, 123rf_filename, description, keywords, country

- Pixta — ID (empty for new), File name, Title (max 100), Keyword, Comment

- Freepik — File name; Title; Keywords (semicolon-delimited file; keywords comma-separated, max 50)

- Archive / All fields — all MetaBrain fields in one file; not for direct platform upload; use for archiving or custom workflows

⚠️ Alamy Image Ref must be filled manually from your Alamy Image Manager account — Alamy assigns this unique ID when you upload your image.

⚠️ Pond5: upload via contributor.pond5.com → Uploads → Apply CSV (not FTP). Select media type first.

⚠️ Dreamstime CSV works for IMAGES only — video CSV returns "must be in JPG format" due to a known Dreamstime platform bug. Enter video metadata manually on their website.

⚠️ Zoonar does not use CSV import — use XMP Sidecars instead (Zoonar reads IPTC metadata embedded in the file).

ℹ️ Getty / iStock: these agencies accept only keywords from their own list, so the file we build for them is adapted to it. A full match is not guaranteed — a few keywords may still come up unrecognised in ESP, and you can simply remove them. We keep improving the match.

Creates a formatted .docx document with one section per image. Useful for agencies that require documents, or for sharing with editors.

Creates a .xmp sidecar file alongside each source image. Read by Adobe Bridge, Lightroom, Capture One, and most DAM software. The recommended method for professional workflows — metadata travels with the file without modifying the original.

Available in the macOS desktop app only.

Permanently embeds metadata (IPTC/XMP) directly into JPEG, PNG, TIFF, and WebP files. For formats that don't support embedding (RAW, EPS, SVG), a sidecar is created instead.

Why it's not in Studio: browsers run in a security sandbox that cannot overwrite files on your disk — they can only offer downloads. This is a fundamental browser restriction, not a missing feature. In Studio, use XMP Sidecars — they carry identical metadata.

⚠️ This option permanently modifies your original files. Keep a backup.

Available in the macOS desktop app only — Export Results → Upload to Stock…

Delivers the finished files straight to the agency, without a browser. Supported: Shutterstock (FTPS), Adobe Stock (SFTP), Pond5, Dreamstime, Depositphotos and 123RF (FTP). Getty and iStock are not here on purpose — they have no self-serve FTP for contributors; you upload in ESP and bring the metadata in with our Getty/iStock CSV.

Your files never move. A file delivered to every agency you ticked gets a green Finder tag Uploaded and stays exactly where it was — so Bridge collections, Lightroom links and the paths in a CSV you already exported all keep working. The tag is searchable in Finder (tag:Uploaded) and you can strip it with a right-click. The dialog also tells you how many of the files in front of you already carry it. A file delivered to only some of the selected agencies is NOT tagged. If you prefer the old behaviour, a separate checkbox moves delivered files into an _UPLOADED subfolder; it is off unless you turn it on. On network volumes and non-Mac filesystems the tag may not be storable — the delivery still goes through, only the mark is skipped.

Embed metadata before upload (on by default) writes the title, description and keywords into the files first, because agencies read the metadata from the file itself over FTP. A file whose embed fails is NOT uploaded — a delivery with empty metadata is worse than no delivery.

Passwords are kept in the macOS Keychain, never in plain text, and are typed once. Three agency quirks worth knowing: Adobe Stock needs a separate SFTP password generated in the contributor portal — generating a new one invalidates the old, so re-enter it here and press Save credentials; Pond5 likewise needs its own FTP password from the dashboard, not your site login; Dreamstime wants your numeric member ID as the username, not your login name.

Test checks both the login and the data channel, because a firewall, a NAT or a TLS-intercepting antivirus can let you log in and still block the transfer itself. If the connection is intercepted, tick "Skip TLS certificate check".

Every successful generation is saved to your account history on the server — metadata plus a small preview thumbnail. History follows your account: process in the App, review in Studio, or the other way round.

Studio → History page; App → Settings → View History. Both show your latest 50 generations. Click any entry to restore the full result to the current session — keywords, title, description, category.

Note: in Mode D, a card is written to history when you press Save in its detail window.

Every MetaBrain account has a unique personal referral link. Share it with a friend. When they sign up and make their first payment, both of you automatically receive +200 credits — no manual steps required.

Open Studio → Settings (gear icon ⚙). Scroll down to the Referral Program card. Your personal link looks like:

`https://metabrain.online/r/MB-XXXXXXXX`

Click Copy link to copy it to your clipboard.

The Referral Program card shows live stats:

- Signups — how many people registered using your link

- Paid — how many of them made their first payment

- Credits won — total bonus credits you have earned

- The +200 bonus is awarded to both parties after the referred user's first payment (any plan or credits pack).

- You cannot refer yourself.

- The referral link is valid for 30 days after a friend clicks it — they must register within that window.

Portfolio Analytics (the "Portfolio Coach") is your personal stock mentor. It looks at your entire published portfolio on Shutterstock, compares it with what the agencies are actually asking for right now — their live creative briefs — and turns that into a concrete action plan: what to fix, what to shoot next, and which themes bring views and sales. The report refreshes every month, so it keeps up with demand. It is a Pro-plan feature and currently in beta.

Nothing to install — no browser extension, no API keys, no "special" links — and it works in any desktop browser (Chrome, Safari, Firefox, Edge).

1. Open metabrain.online/studio → Portfolio Analytics and press Generate my button.

2. Drag the "MetaBrain Sync" button to your browser's bookmarks bar (press ⌘⇧B / Ctrl+Shift+B if the bar is hidden).

3. Open your contributor dashboard at submit.shutterstock.com (you are already logged in) and click the bookmark.

4. It collects your catalog and your all-time sales automatically — a small counter shows progress — then come back to the Analytics page and the report builds itself.

Click the bookmark again anytime to refresh. Until you click it, nothing is collected or stored.

Once processed, the coach report includes:

- KPIs — lifetime revenue, downloads, revenue-per-download (RPD), dead-weight share, and average keywords per file.

- Charts — revenue by month, uploads by month, format mix (photo / video / illustration / vector), and selling vs dead weight.

- Agency brief coverage — how much of each live agency brief your keywords already cover, with links to the sources.

- Demand gaps — themes the agencies want that your portfolio is missing, each with a generated example shot.

- Your shot plan — concrete shooting assignments matched to demand, with example titles and keywords, exportable to PDF.

- Files to fix first — your weakest files (thin keywords, short descriptions) with a one-click "Fix in Studio".

- The coach verdict — a plain-language summary plus a monthly quota (new uploads, video clips, files to re-keyword).

Yes — here is exactly how it works:

- The button runs on your own Shutterstock page under your own session, as if you opened your dashboard yourself. It never sees or sends your password.

- Only aggregated numbers leave your browser: counts, keyword frequencies, monthly totals, per-file sales summaries. Never your images, your catalog files, or full statements.

- That data is stored on MetaBrain's own servers (where your account already lives), sent over HTTPS, tied only to you, and used only to build your report. We never sell it, share it, or train models on it.

- You stay in control: "Disconnect portfolio" deletes all collected data and reports from MetaBrain at any time. Your Shutterstock portfolio itself is never touched.

- The button's code is open and self-contained — you can read all of it in the bookmark; nothing is hidden and nothing loads from elsewhere.

The coach re-checks your portfolio automatically every month. You can also press Re-scan now on the report (once per day), or just click the bookmark again to push fresh data. Recommendations are built from public agency briefs and your portfolio composition — they improve your odds of being found, but no income is promised. Portfolio Analytics is included in the Pro plan and is currently in beta.

Processing time varies with AI provider load and content complexity. Images typically take 5–20 seconds; videos take longer — frames must be extracted and a larger payload analysed, so long clips can take a few minutes. Transient errors are retried automatically.

Not directly. Convert RAW to JPEG or TIFF first. Lightroom users: export a JPEG, process it in MetaBrain, then import the XMP sidecar — Lightroom applies the metadata to the original RAW.

Images: JPEG, PNG, TIFF, WebP

Vectors: EPS, AI, SVG — in both the App and Studio (Studio converts EPS/AI on our server automatically; the desktop App needs Ghostscript installed locally)

Video: MP4, MOV everywhere; AVI, MKV, WebM in the App; WMV, M4V, MTS, MXF additionally via drag-and-drop in Studio

HEIC is not supported — convert iPhone shots to JPEG first.

No browser decodes professional codecs such as ProRes; Studio names the exact codec in the error message. Drop the H.264 proxy your editor already writes and set "Master file name" in export so the CSV names the master, or transcode a copy with Shutter Encoder or HandBrake (free on Windows, macOS and Linux).

Files are processed in memory and never stored permanently. What we keep is the generated metadata (title, description, keywords) and a small preview thumbnail in your history — so you can restore past results.

By default the keyword count is a quality ceiling, not a quota: semantically duplicate terms are removed^*, and narrow subjects simply have fewer genuinely relevant terms. If you need an exact count, enable the strict-count option in Output Settings.

Plan credits reset monthly and do not roll over. Credits Pack credits (+500 one-time purchase) never expire.

Settings → Cancel subscription, confirm — done. Cancellation is immediate and self-service; your plan stays active until the end of the paid period, then the account reverts to Free. Your data, history, and remaining pack credits are kept.