Every studio as tools

Get credit balance

Return the signed-in Lumen Brain user's credit balance, plan, monthly renewal date and Brain storage (bytes used against the room the plan gives, plus any storage an admin gifted on top). Free to call.

List generations

List the signed-in user's recent Lumen Brain generations with model, credits spent, a 24-hour signed download URL and, while the output's link is open (it is by default), its permanent public URL — the one to paste into other tools or docs. Every item carries its generation id: pass ids to save_output_to_brain to file outputs in a Brain folder, or as reference_generation_ids / reference_video_generation_id to render from them again. Renders still on their way are named first with their job ids (wait_for_generations returns them once done), so the newest finished item is never taken for one still rendering. search finds an earlier output by words of its prompt. Free to call.

Generate image

Start an image in Lumen Brain and return as soon as the job is running — the render goes on in the background, so the conversation is free again at once. The inline preview follows the job and shows the picture the moment it lands (usually well within a minute); it is saved to the gallery either way, and a failed render refunds its credits. When a next step in the same task needs the finished picture — placing it in a design, animating it, filing it, reviewing it — call wait_for_generations with the returned job_id: it returns the generation id and links; otherwise do not wait. Models and credit costs: lite (Nano Banana 2 Lite, 1), fast (Nano Banana 2, 1), nano (Nano Banana, 2), pro (Nano Banana Pro, 4), gptImage2 (GPT Image 2, 6), flare (GPT Image 2.5 Flare, 6 — OpenAI's fast default, about twice the speed of GPT Image 2), sunburst (GPT Image 2.5 Sunburst, 8 — OpenAI's most precise, for detailed edits and finished campaign work), qwen (Qwen Image, 2 — an open model on Replicate with its safety checks off, for pictures the others refuse; 1K only), qwen3 (Qwen Image 3, 2 — Alibaba's newest, the sharpest text and layouts, 1K or 2K; hosted by Alibaba, so its content rules apply), seedream (Seedream 5.0 Pro, 2 — ByteDance's newest: it thinks a long brief through, holds a dense layout together and writes words into the picture in fourteen languages, 1K or 2K), seedreamLite (Seedream 5.0 Lite, 1 — the same family, quicker and a third of the price), soul2 (SOUL 2, 1 — Higgsfield's photographic people model), grok (Grok Image 2.0 by xAI, 2 — renders and edits), recraft (Recraft V4.1 Pro, 6 — typography and design, 1K or 2K), and ideogram and ideogramHigh (Ideogram 4.5, 2 and 6 — posters, logos and lettering, 1K or 2K, at its medium and high quality; it renders, and edits the first reference picture with up to four more as references); 2K output costs 1.75x and 4K 3x the base credits, rounded up. Reference images (inline base64 and/or media-library assets) drive image-to-image work — edits, restyles, or keeping a face or product consistent — up to the model's own cap: lite, nano, qwen and qwen3 take 3, grok 4, ideogram and ideogramHigh 5 (the first is the picture edited), fast 6, the Seedream pair 10, pro 14, gptImage2, flare and sunburst 16; soul2 and recraft take none. Large Brain pictures are sent as a 2048 px copy so the references fit one job. An earlier output goes in the same way — "with this one, now…" — as reference_generation_ids (or reference_latest_output for the newest), read straight from the gallery. For SEVERAL images (a batch, variations, N of something) call queue_generations instead of repeating this tool. For a styled look, pass a curated preset name (see list_presets) as preset — the platform applies the style server-side at no extra credit cost. If the prompt refers to a picture the user attached to the chat and no reference reached the call, nothing is rendered: the tool opens the upload widget in the chat instead (no credits spent) — tell the user to drop the file in, then call get_upload_session and run again with reference_asset_ids.

Generate video

Queue a short video clip (Gemini Omni Flash 1.1, Gemini Omni Flash, MiniMax Hailuo H3, Kling 3.0, Kling 2.6 Pro, Kling 2.5 Turbo Pro, PixVerse 6, Seedance 2.0, Seedance 2.5, Wan 3.0, Genjutsu or Cinema Studio 4.0) in Lumen Brain. Returns at once with a job id: the clip renders in the background, the inline preview shows it the moment it lands, and it is saved to the user's gallery — call list_generations later for its signed URL. Credits are reserved on queueing and refunded automatically if the render fails. Several clips? Call this once per clip, back to back — they render in parallel, each with its own preview — or use queue_generations for a large batch of one prompt. Costs 1-8 credits per second on Gemini Omni Flash 1.1 — the cheapest model here — 3 on Gemini Omni Flash, 3-4 on Hailuo H3, 4-13 on Kling 3.0, 5 on Kling 2.6 Pro, 3 on Kling 2.5 Turbo Pro, 2-5 on PixVerse 6, 4-9 on Seedance 2.0, 6-13 on Seedance 2.5, 2-7 on Wan 3.0, 6-13 on Cinema Studio 4.0, and on Genjutsu 3-4 on Hailuo H3 or 1-8 on Gemini Omni, depending on resolution (e.g. 6s = 6-78). GENJUTSU is motion transfer: a clip goes in (as reference_video_url / reference_video_asset_id / reference_video_generation_id — it renders from nothing else) and the same motion, camera and timing come out, recast with whoever or whatever takes over. It renders on one of two engines. Key genjutsu runs on Hailuo H3: a clip of 5-15s and up to 9 cast pictures through reference_asset_ids — the people, the product, the clothes — named @Image1, @Image2 … in the prompt ("@Image1 replaces the dancer"); once a picture is passed the prompt is optional; 768p at 3 or 2K at 4 credits/s. Key genjutsuOmni runs on Gemini Omni, for less: a clip of 3-9s under 32 MB and up to 9 cast pictures the same way — Omni reads pictures and a clip together on its reference endpoint, so the clip is cut on the server into up to three consecutive parts of under 3s; once a picture is passed the prompt is optional, and without pictures the prompt is required and says who or what takes over ("a chrome robot replaces the dancer"); 360p at 1, 720p at 3, 1080p at 4, 4k at 8 credits/s. On either engine the output runs as long as the clip, so its length is read from the clip, and the duration parameter is only used for a clip given by URL, or a Brain or gallery clip whose length is not recorded — pass the clip's own length then. CINEMA STUDIO 4.0 (key cinema4) is Seedance 2.5 with Higgsfield's direction on top — 4-30s, 480p or 720p, up to 30 reference pictures, a reference clip and sound — and takes direction, the Director's Panel as fields: genre, era, pacing, light, camera_model, camera_lens, camera_aperture, camera_movement, color_palette, each optional (left out, the model reads the prompt for it). A reference clip on Cinema Studio is billed too: Higgsfield meters the clip's seconds together with the render's at 0.6× the rate, so the price covers both. Optionally animate a reference image, either inline base64 or a media-library asset. Seedance 2.0, Seedance 2.5, Hailuo H3 and Gemini Omni Flash 1.1 read a whole SET instead — up to 9 pictures (Wan 3.0 up to 4) through reference_asset_ids (a character, a second character, the place they are in), arriving as @Image1, @Image2 … for the prompt to name ("the character from @Image1, in the room from @Image2"); on Seedance a set caps the output at 720p the way a clip does. On Omni Flash 1.1 one picture alone is the starting frame, while two or more — or any beside a reference clip — go to its reference endpoint as a set. Gemini Omni Flash 1.1, Hailuo H3, both Seedance generations and Wan 3.0 also take a reference VIDEO by public URL (reference_video_url): Omni Flash 1.1 carries the clip on or edits it — that is how a scene grows past 10s, 10s at a time to Google's 40s ceiling — and with pictures beside it reads the clip and the pictures together: its reference endpoint takes the clip's last 10s, cut on the server into up to three consecutive parts under 3s, and renders a new shot of the chosen 3-10s; Hailuo, Seedance and Wan use the clip as a reference to continue, restyle or borrow motion from; a clip caps the output at 720p on Seedance and Wan. Both Seedance generations also take SOUND — up to 3 MP3/WAV files by public URL (reference_audio_urls, 15s and 15 MB each), which arrive as @Audio1, @Audio2 …: say in the prompt what each is for ("BGM references @Audio1", "the voice references @Audio1"). Sound needs a reference image or clip with it, and caps the output at 720p just as a clip does. A clip already in the Brain (uploaded through media_upload_widget, or listed by list_media / search_brain with kind "video") goes in as reference_video_asset_id instead of a URL. For a styled look, pass a curated preset name (see list_presets) as preset — the platform applies the style server-side at no extra credit cost. If the prompt refers to a picture or clip the user attached to the chat and no reference reached the call, nothing is queued: the tool opens the upload widget in the chat instead (no credits reserved) — tell the user to drop the file in, then call get_upload_session and run again with reference_asset_ids or reference_video_asset_id.

Wait for generations

Wait for renders that generate_image, generate_video or queue_generations started, and get their outputs: for each finished job its generation id (what create_design, update_design, save_output_to_brain and the generate tools' reference_generation_ids take), a 24-hour signed URL and, while the output's link is open, its permanent public link. The generate tools return as soon as a job is running, so the conversation stays free — call this only when a next step in the same task needs the finished output (placing a picture in a design, animating it, filing it, reviewing it). Waits up to timeout_seconds (default 45) for all of them; any still rendering by then is reported as such — call again to keep waiting. A failed render is reported with its reason; its credits are refunded. Free to call.

Queue generations

Queue one or many generations to render in the background instead of waiting for each one. ALWAYS use this instead of repeated generate_image / generate_video calls when the user asks for several images or clips (a batch, variations, N of something): every job is enqueued at once, credits are reserved up-front and refunded on failure, the studios show live progress tiles, and results land in the gallery — check list_generations a few minutes later. Same models, credit costs, reference-image rules and preset support as the generate tools; video jobs also take a Brain clip as reference_video_asset_id for the video-to-video models. Video jobs on Seedance 2.0, Seedance 2.5, Hailuo H3 and Gemini Omni Flash 1.1 read up to 9 reference pictures at once (Wan 3.0 up to 4) (a character, a second character, the place they are in), arriving as @Image1, @Image2 … for the prompt to name; every other video model animates the first as its starting frame, and so does Omni Flash 1.1 given one picture and no clip. Earlier outputs go in too: reference_generation_ids (images), reference_video_generation_id (a clip, video jobs) or reference_latest_output for the newest one — read from the gallery server-side, nothing re-uploaded. Returns the queued job ids immediately. If the prompt refers to a picture or clip the user attached to the chat and no reference reached the call, nothing is queued: the tool opens the upload widget in the chat instead — tell the user to drop the file in, then call get_upload_session and run again with reference_asset_ids (or reference_video_asset_id).

List curated presets

List Lumen Brain's built-in curated presets — style packs maintained by the platform (film noir, product studio, anime, analog film, UGC, cyberpunk and more). Each has a name, a title and a short description of the look it produces. To use one, pass its name as the preset parameter of generate_image or generate_video: the platform then rewrites the prompt to match the style and attaches the preset's reference images server-side. The style brief itself is not exposed — only the preset name is needed. Check this before generation work when the user asks for a styled or on-brand look. Free to call.

Media upload widget

Open an in-chat upload widget so the user can pick local images (PNG/JPG/WEBP), video clips (MP4/MOV/WEBM) or sound files (MP3/WAV), up to 100 MB each and 25 MB for sound, from their device and drop them straight into a media library folder. Call it IMMEDIATELY, without asking first, whenever the user's message refers to a picture, clip or file they attached to the chat, pasted, or have on their device ("the photo I attached", "this image", "my logo", "csatolt kép") — chat attachments never reach these tools, and asking the user to request an upload is a wasted turn. Make it the only tool call of that turn and tell the user to drop the files in. The widget uploads directly from the chat; a browser fallback link is included. For a file that is already online (a share link, a CDN file, a Lumen output link) call save_media_from_url instead — no widget needed. Returns a session id; after the user uploads (hosts that support it let the widget message you itself), call get_upload_session to see the stored assets — image ids go into generate_image / queue_generations as reference_asset_ids, a clip's id into generate_video as reference_video_asset_id, and a sound file's into generate_video as reference_audio_asset_ids.

Get upload session

Check an upload session started with media_upload_widget: returns the status and the assets the user uploaded (id, name, folder) so you can reference them in prompts or with get_media.

List media folders

List the signed-in user's media library folders (Drive-style tree with full paths, e.g. /Brand/Logos) with asset counts and folder notes. Start here before prompting from a brand folder. Free to call.

List media assets

List the image, video and sound assets stored in the user's media library, optionally inside one folder path (e.g. /Brand/Logos) and optionally including nested subfolders. Returns the folder's own note first — treat it as authoritative context for any task about this folder — plus asset names, kinds, notes and tags, not the bytes. Use get_media to read an image (or a clip's poster frame and signed URL); a clip's id goes into generate_video / queue_generations as reference_video_asset_id, and a sound file's (kind "audio") into generate_video as reference_audio_asset_ids. Free to call.

Search the Brain

Search the user's whole Brain (media library) in one call — every folder at once — matching asset names, notes and exact tags, images and video clips alike. Each hit includes its folder path and kind, so you can follow up with get_media for the bytes (a clip's poster frame and signed URL) or list_media to browse around it; a clip's id goes into generate_video as reference_video_asset_id. Free to call.

Get media asset

Read one stored asset from the media library by asset id or by folder path + name. An image comes back as base64 image content to use as a visual reference when prompting; a video clip comes back as its poster frame plus a 24-hour signed URL, and its id goes into generate_video / queue_generations as reference_video_asset_id. Free to call.

Create media folder

Create a folder (and any missing parents) in the media library, e.g. "/Brand/Logos". Optionally attach a note describing how the folder should be used when prompting.

Save media asset

Save a base64 image into the user's media library, in a folder path (created if missing). Use it to build brand folders that later prompts and bulk runs can read. Images only — a video clip reaches the Brain through media_upload_widget. For a file that already has a public URL, call save_media_from_url instead and skip the base64.

Save media from a public link

Import a picture or a video clip that already lives at a public http(s) link — a share link, a CDN file, an image on a website, or a Lumen Brain output link (lumenbrain.app/o/…) — straight into the user's media library, in a folder path (created if missing; defaults to /Uploads). The server fetches the file, so no base64 travels through the chat: prefer this over save_media whenever the file has a URL, and over media_upload_widget whenever the file is online rather than on the user's device. Images (PNG/JPG/WEBP/GIF/AVIF) and clips (MP4/MOV/WEBM), each up to 100 MB; a big image keeps its original in storage with a working copy inline. Up to 10 links per call; each becomes its own asset, and the returned ids go into generate_image / queue_generations as reference_asset_ids or into generate_video as reference_video_asset_id. Filing spends no credits.

Save output to Brain

File finished generations into a Brain folder (created if missing) — the way to put outputs where the user wants them, done server-side, with nothing downloaded or re-uploaded. Pass generation ids (a generate tool's result carries generation_id; list_generations lists them) or latest for the newest N finished outputs. An image becomes a Brain image in the folder (a big one keeps its original in storage with a working copy inline); a clip becomes a Brain clip, ready as a video input. Filing spends no credits and leaves the gallery as it is. To file an output the moment it renders, pass brain_folder to generate_image, generate_video or queue_generations instead; to rearrange later, use move_media.

Move media asset

Move a Brain asset (an image or a clip) into another folder, optionally renaming it — for arranging the library: sorting outputs and uploads into the folders they belong to. The target folder is created if missing; the asset keeps its id, tags and links. Find the asset by id (list_media, search_brain, a filing result) or by its current folder path and name.

Set folder note

Write or update a folder's note in the media library. The note is the folder's persistent context — usage rules, brand guidance, learnings from past work — read back by list_folders and list_media in every future session, so record here anything worth remembering about the folder and its contents. mode=replace rewrites the whole note (an empty string clears it); mode=append keeps the existing note and adds this text on a new line. Only the folder's owner can write. Free to call.

List custom skills

List the user's stored custom skills — reusable instruction packs with a name, title and a "when to use" description — including skills other people shared with them, marked with who shared each. Check this before generation or brand work: when a skill's description matches the task at hand, load it with use_skill and follow it. Free to call.

Use a custom skill

Load one stored skill's full instructions and activate them for the current task. Once loaded, the skill governs how you work in this session: follow it when writing generation prompts, when choosing models and settings on generate_image/generate_video, and when persisting context (set_folder_note, asset notes). Free to call.

Save a custom skill

Store or update a custom skill so any future session can load it with use_skill. Use this when the user asks to save a workflow, style or set of rules as a skill. Write the description as a trigger ("when to use this") and the instructions as directives covering prompt style, model/settings choices and what context to persist. Saving under an existing name overwrites that skill.

List design templates

The Design Studio's templates: the presets everyone has (the built-in ones referenced as starter:<id>, the ones an admin shares by id) and the user's own saved templates (by id), each with its page size, tags, slots — the bound layers a template fills in: text slots (headline, body, cta…) with the most characters each takes (max_chars) and image slots (image, portrait, logo…) — and its own instructions, the template's system prompt for whoever writes into it: follow it when writing values for create_design. Pass a reference to create_design (with a value per slot) or generate_design_variants; get_design shows a template's slots again with include_doc for its full document. Free to call.

List designs

The user's saved designs and templates in the Design Studio, newest first: id, title, kind, page size, tags. get_design shows one's slots and document; render_design turns one into a picture. Free to call.

Get a design

One design or template by reference (a design id, or a starter as starter:<id>): its page size, tags and slots; with include_layers a table of its layers — id, name, type, box, what each holds — the ids edit_design takes; with include_doc the complete design document, the JSON the studio edits, to read or to change and pass back to create_design / update_design as doc. Free to call.

Create a design

Saves a new design (or template) in the user's Design Studio: from a template reference with a value per slot — text as it is; an image slot takes a Brain asset id, a gallery generation id (an output of generate_image / queue_generations), an https link, or "empty" — or from a complete design document you composed (doc, in the shape get_design returns with include_doc). Returns the saved design with its slots and an editor link; with render: true also queues a picture of it (see render_design). Free; the AI tools are the ones that spend credits.

Update a design

Changes one of the user's designs or templates: pour new values into its slots (text, or a picture as a Brain asset id / gallery generation id / https link / "empty"), replace its document, rename it, make it a template or a design again, retag it. Nothing else changes. With render: true also queues a picture of the result.

Edit a design's layers

The editor's own moves on one of the user's designs, as a list of operations applied in order: add_text, add_picture, add_frame, add_shape, add_gradient; set (any property of a layer — words, font, size, colour, box, slot key, max_chars, a frame's picture or hint, a shape's fill), move, resize, remove, duplicate, reorder, align; set_background (a colour, a gradient or a picture), remove_background_picture, resize_page, set_instructions (the template's system prompt), fill_frames (pictures into the empty frames). Layers are named by id, name or slot key — get_design with include_layers lists them. Every operation is checked against the document's rules and the whole is saved once; the layers come back as a table, with a warning when copy prints over copy or runs off the page. Free; with render: true a picture is queued.

Check a design's layout

Measures a design the way the Studio does before a post goes out: which text prints over other text, and which layers run off the page. With repair: true the layout is put right the way the Studio's own fit does it — air under grown copy closes first, what sits below moves down, the stack slides up when the page is short, and only then is the type set smaller (to 75% of its size; strict: true allows 60%) — and the repaired design is saved when it is one of the user's own. Free.

Delete a design

Deletes one of the user's designs or templates for good. Pictures already rendered from it stay in the gallery.

Remove a picture's background

The subject cut out of a picture and everything else made transparent: a packshot off its studio floor, a person off their room, a logo off its card. BiRefNet's matting keeps hair, fur and soft edges where a plain mask smears them. Point at a picture in the Brain (asset_id), an output in the gallery (generation_id) or a public link (url). The cut-out is filed as a transparent PNG of its own, beside the original in the same folder, and the original is left as it is. The returned asset id goes straight into create_design, update_design or edit_design as a picture. Free, within the day's allowance the user's plan carries; the answer says how many are left. If hair or fur comes back rough, call again with profile: "fine".

Share a design with everyone (admin)

For an administrator: makes one of their designs or templates a preset everyone has — it appears in every user's Design Studio and in list_design_templates, usable by all and changed by its owner alone — or takes it back. A shared template that started from a built-in starter stands in for that starter. Anyone else is refused.

List design page sizes

The page sizes the Design Studio offers, by group (Social, Ads, Print, Other): each with its id — what generate_design, create_design's documents and edit_design's resize_page take as preset — its name, its pixels and its ratio. Any width × height in pixels is accepted too; these are the formats people actually post, print or pin. Free, needs no sign-in.

List design fonts

The typefaces a design can name: the built-in families (Google Fonts, by category — sans, serif, display, handwriting — with the weights each ships and whether it has an italic) and the fonts the user uploaded themselves in the Studio, by family name. A layer's font in edit_design and a document's fontFamily take any family listed here; brand fonts passed to generate_design should be from here too. Free to call.

Generate a design with AI

The AI Post Editor: from a brief — the message, the audience, the tone, the words to use — a complete design is written and saved (layers of copy, colour and picture slots; every piece of copy and every image slot bound as a slot, so it works as a template). Size by preset (ig-square, ig-portrait, ig-story, fb-post, linkedin-post, x-post, yt-thumb, pinterest, fb-cover, linkedin-banner, ad-square, ad-landscape, ad-leaderboard, ad-mrec, a4-portrait, a4-landscape, a5-flyer, business-card, presentation, og-image, email-header, wallpaper) or width × height in px. Optionally restyle an existing template or design for the brief (from), and pass brand colours, fonts and voice. Models: sonnet (Claude Sonnet 5, 2 credits) or opus (Claude Opus 5, 5 credits). Returns the saved design with its slots; picture slots come back empty — fill them with update_design (a Brain asset or a generate_image output) and render_design for a picture, or pass render: true here to queue one at once. conversation: true also opens the brief as a conversation on the AI Post Editor's Create tab, where the user (or continue_design_conversation) takes it further turn by turn.

Revise a design with AI

One instruction applied to one of the user's designs by the AI editor — "make the headline shorter and bolder", "swap to a dark palette", "add a price tag bottom right" — and the whole document comes back revised and saved in place, ids and slots kept. sonnet costs 2 credits, opus 5. With render: true also queues a picture of the result.

Generate variants of a template with AI

Bulk content from one template: the AI writes the copy for N versions — from a brief (each variant a different angle), or one variant per line of rows (a pasted spreadsheet column: a product, a quote, a city…) — and each set of values is poured into a copy of the template and saved as its own design. Picture slots keep the template's picture; for each variant the AI's picture prompt is returned, so pictures can be made with queue_generations and placed with update_design. Costs 1 credit per started ten variants on sonnet, 2 on opus. With render: true a picture of every variant is queued. batch: true also writes the run to the AI Post Editor's Bulk tab as a batch (list_bulk_batches), where the user can review, edit and render the posts; project: true gathers the saved variants into one design project (list_design_projects), a document the Studio opens page under page.

Research a topic on the web

The web read the way the AI Post Editor reads it before writing about something current: a few focused searches planned from the brief, the answers merged, kept to what is about the subject and recent, and returned as dated facts with their sources — names, numbers, dates, prices, results as they stand today. Use it before writing copy about this week, a launch, a price, a result, an event, or anything else that has a date; pass rows (one line per post of a batch) and the plan covers the batch. 1 credit when something current comes back; nothing when the web has nothing new to say. generate_design and generate_design_variants run the same search themselves with research: true.

Find pictures on the web

Pictures for a few words, found the way the Design Studio fills a frame: the free stock libraries (Pexels, Unsplash, Pixabay, Openverse, Wikimedia Commons) first, and when they have nothing, an engine (Gemini with Google Search, or Claude) searches the open web for a page showing a public picture; the best matches are fetched and filed into the Brain — under /Design Studio/Web pictures unless a folder is given — with the photographer and the licence as the note, so each comes back as a Brain asset id ready for a picture slot (create_design, update_design, edit_design's fill_frames). Free from a stock library; 1 credit for an engine's search that found something. Name the real subject — the product, the person, the place, the event — not a mood.

Review finished designs with AI

The art director's check the Studio makes before a post goes out: Claude looks at finished renders (get_design_render must say done) and says, for each, whether it can go out — naming text printing over text, anything cut off by the page edge, copy that cannot be read where it sits, a badge or headline covering the subject of the photograph, an empty frame or a placeholder left in. Taste and tone are not judged. Up to 8 renders a call; 1 credit per started ten. A post that fails is put right with edit_design, check_design_layout (repair: true) or revise_design, then rendered and reviewed again.

Render a design

Turns a design into a picture, saved to the user's gallery with a public link. Pass a design or template reference (optionally with values for its slots, so a template renders filled without being saved first) or a complete document. Painting takes a browser: where this chat shows Lumen Brain views the picture is painted right here in the result view; otherwise it is painted the moment the user has Lumen Brain open in a browser (the studio paints waiting renders itself), or by opening the returned page_url once. Call get_design_render with the render_id for the picture's links when it is done. Free.

Get a design render

Where a render stands, by its render_id: pending (not painted yet — the user's open Lumen Brain or the page_url paints it), failed (with why), or done with the picture's links — a public link that never expires and a signed one — and, when it can be fetched, a small preview as image content. Free to call.

Render every page of a project

A picture of every page of one of the user's projects, in order — one render each, saved to the gallery with public links as render_design does. Painting takes a browser: where this chat shows Lumen Brain views the pages are painted right here; otherwise the moment the user has Lumen Brain open, or by opening a page_url once. get_design_render reads each render by id. Free.

List design projects

The user's design projects — a batch of posts kept as one document with pages, the way a Canva design has pages; the Studio opens a project and lays its pages out one under the other. Newest first, with how many pages each holds; with project one project in full, its pages in order (each a design id get_design, edit_design and render_design take). Free to call.

Create a design project

A new multi-page project in the user's Design Studio: each page either a copy of a template or design with values poured into its slots (from + values), or one of the user's own designs moved in as a page (design). Pages keep the order given. Returns the project with its page ids — each a design that get_design, edit_design and render_design take — and its editor link; render_design_project paints every page. Free.

Update a design project

Changes one of the user's projects: rename it; add pages at the end (a copy of a template with values, or one of the user's designs moved in); take pages out — detach_pages makes them standalone designs again, remove_pages deletes them; and set the order of every page with order (the page ids, first to last). Free.

Delete a design project

Removes one of the user's projects. Its pages go with it, unless keep_pages: true, which first makes every page a standalone design in the library. Cannot be undone.

List bulk batches

The batches on the user's AI Post Editor Bulk tab — every run of many posts from one template, saved from its first keystroke: the template, the brief, the pasted rows, the model and the switches, and with batch one batch in full: every post's values, its picture prompts, whether it is selected, and what the review said of it. generate_design_variants with batch: true writes a run here so it appears on the tab. Free to call.

List AI editor conversations

The briefs begun on the user's AI Post Editor Create tab, each kept as a conversation: the brief, the turns (what was asked, what the editor answered) and the design as it stands. Newest first; with conversation one in full — its turns, and its design's slots and layers. continue_design_conversation takes one further; generate_design with conversation: true starts one. Free to call.

Continue an AI editor conversation

One more turn on a conversation from the user's AI Post Editor: the instruction is applied to the design as it stands ("make the headline shorter", "swap to a dark palette", "add a price tag bottom right"), the revised design becomes the conversation's current one, and both turns are kept — the tab shows them when the user opens it. sonnet costs 2 credits, opus 5. With save: true the result is also saved as a design of its own (render: true then queues its picture).

Delete an AI editor conversation

Removes a conversation from the user's AI Post Editor — its turns and the design it held. Designs saved to the library from it stay. Cannot be undone.

List voice models and voices

The voice studio's catalogue: the text-to-speech models — Eleven v4 (the newest and the default), Eleven v4 Turbo (cheaper, near-instant) and Eleven v3, all ElevenLabs through fal — with their price per started 1,000 characters, and every preset voice with a line on how it sounds and the words it is found by (old, deep, warm, young, british, conversational…). search finds a voice by any of those. Signed in, it also lists the user's own voices (cloned from a recording, designed from a description, or added from ElevenLabs' Voice Library by ID — pass them to generate_speech as their custom:<id> key; a voice cloned or designed on MiniMax is read by MiniMax Speech 2.8 HD) and marks the voices they hearted with favorite: true; favorites_only narrows the list to those. Pass a model key and a voice id to generate_speech; voice_preview plays a sample; set_voice_favorite hearts one. Free.

Hear a voice

A sample of one voice reading a fixed line — made once and kept, so it comes back at once for a voice anyone has heard before. Returns an MP3 link to play or pass on: public for a preset, signed for an hour for one of the user's own voices ("custom:<id>", whose sample is the line it read when it was made). Free: the script never changes and only the catalogue's voices are read.

Generate speech (text to speech)

A script read aloud and saved to the user's gallery as an MP3 with a public link. Models (ElevenLabs, through fal): elevenV4 — Eleven v4, the newest and the default, the most natural read in 90+ languages; elevenV4Turbo — the same voices for less and nearly at once; elevenV3 — the previous model, for voices tuned on it. Voices: a preset from list_voices, or one of the user's own voices as "custom:<id>" (list_voices lists them when signed in) — a voice cloned or designed on MiniMax is read by MiniMax Speech 2.8 HD whatever model is named. Direction goes in square brackets in the text, in English whatever language the script is in: a delivery tag right before the words it colours ([warmly], [whispers], [excited], [sarcastic]) and a reaction where it happens ([laughs], [sighs], [gasps]) — about one per sentence or two; the brackets are never read out. Pronunciation goes in slashes as IPA. Optional stability (0 most expressive … 1 most even, default 0.5), similarity (v4 models only, default 0.75) and language (ISO 639-1; left out, it is read from the script). Brackets, stability and similarity are ElevenLabs' — a MiniMax voice takes only language, and its brackets are dropped. Up to 5,000 characters a call on v4 and Turbo, 6,000 on v3; longer scripts go in parts. Priced per started 1,000 characters at the model's rate (list_voices says it). Returns the file's links and the generation id — usable as reference audio for the video models that take a voice track, and filed into a Brain folder with save_output_to_brain.

Heart a voice

Adds a voice to the user's favourites, or takes it off (favorite: false). Favourites lead the voice studio's lists and list_voices marks them (favorites_only narrows to them) — a way to keep the few voices a brand uses at hand. Takes a preset voice id or one of the user's own as "custom:<id>". Free.

Add an ElevenLabs voice by ID

Adds any voice from ElevenLabs' Voice Library (elevenlabs.io → Voice Library → a voice → its voice ID, twenty letters and digits) to the user's own voices. It reads one sample line first (on Eleven v4, or v3 when v4 does not take it) — a voice fal cannot read is never added — and from then on it is "custom:<id>" in list_voices, generate_speech, voice_preview and change_video_voice. 1 credit, refunded if the voice does not work.

Clone a voice from a recording

Makes a voice of the user's own from one recording of a person speaking — their own voice, or one they have the speaker's permission for — and keeps it among their voices as "custom:<id>" for generate_speech, voice_preview and set_voice_favorite. ElevenLabs clones it (through fal), so it reads on Eleven v4, v4 Turbo and v3 and can re-voice a video with change_video_voice; should that be unavailable, MiniMax clones it instead and MiniMax Speech 2.8 HD reads it (not usable for change_video_voice). The recording: a sound file in the Brain (asset_id), a public https link (audio_url), or the bytes (audio_base64, mp3/wav/m4a), 10 seconds to 5 minutes of clear speech from one speaker — a minute is plenty — at most 20 MB. BEFORE calling, ask the user to confirm: "This is my own voice, or I have the speaker's explicit permission to clone it. I won't use it to impersonate anyone or to mislead." — and pass consent: true only when they have. 40 credits, refunded if the clone fails. Takes up to a minute.

Design a new voice from a description

Makes a brand-new voice — nobody's, made up from the words — from a description of how it should sound: age, gender, accent, character, pace, what it is for ("a deep, calm male narrator in his sixties, warm and unhurried, for documentaries"). MiniMax designs it and MiniMax Speech 2.8 HD reads it (Hungarian included); it is kept among the user's voices as "custom:<id>" for generate_speech, voice_preview and set_voice_favorite (not for change_video_voice, which takes ElevenLabs voices). It reads a first line — in its language where one is written (hu, en, de, fr, es, it), in English otherwise — returned as a sample link. 80 credits, refunded if it fails.

Change the voice in a video

Re-voices a video with ElevenLabs and puts the new voice back on the picture — the Higgsfield-style voice change. The speech is lifted out, ElevenLabs' voice changer reads it in the chosen voice (timing, breath and emotion kept, so it stays on the lips), and with keep_background (the default) the music and effects are mixed back under it. The source is one of: generation_id (a finished clip in the gallery — any generate_video output), asset_id (a video in the Brain) or video_url (a public https link). The voice is a preset from list_voices or one of the user's own ElevenLabs voices as "custom:<id>" (not one cloned or designed on MiniMax). Up to 5 minutes and 50 MB; every speaker gets the same new voice. 12 credits a minute (1 per started 5 seconds, at least 2), refunded if it fails. Returns at once with an id — the change runs in the background (a minute or two) and lands in the gallery as a new clip; get_voice_change follows it.

Follow a video voice change

Where a change_video_voice run stands — and each call moves it on: the step it is on, then the finished clip's gallery id and a signed link, or why it failed (credits are refunded then). A change takes a minute or two; there is no need to call this in a tight loop.

Transcribe audio (speech to text)

A recording turned into text — a voice memo, an interview, a clip's soundtrack, a voiceover made with generate_speech — by the transcription model, and the transcript saved to the user's gallery. Point at a gallery output by generation_id (a tts output, or a clip), a public url, or send the bytes as audio_base64 (mp3, m4a, wav, webm, ogg, flac; up to 25 MB). 1 credit per started five minutes; when the recording's length is not given it is taken from its size at 128 kbit/s, so pass duration_seconds when it is known.