The Studio

Pictures, clips, music, speech.
On your key, into your folder.

Image, video, music and speech from any model your OpenRouter key reaches. The panel reshapes itself around the model you pick, and every result is a file in the project with its prompt and price beside it.

VectorBrain’s Studio. On the left, the Image tab with a prompt for a lighthouse at first light, the Gemini 2.5 Flash Image model at $0.039 an image, and Versus switched on against GPT Image 1. On the right, the stage shows both results side by side, each with a button to pick it as the winner.

VectorBrain’s real Studio, running on example data. The harbor and lighthouse pictures are procedural stand-ins drawn for these captures, not model output.

01

The panel is drawn by the model in it.

Resolution, duration, aspect, start and end frames, audio and seed appear only when the chosen model publishes them, so no control on screen does nothing. A model that publishes nothing gets generic controls, because unknown is not the same as none.

Where the model list comes from
The Studio’s Video tab with two models. With Veo 3.1 at $0.40 a second, the panel offers resolution, length, aspect, start frame, end frame, references, audio and a seed. With a text-only example model at $0.05 a second, only length, aspect and references remain.
02

Two models, one prompt, side by side.

Switch on Versus and Generate sends the same prompt to both models at once. Neither is shown until both land, so neither is judged first. Your picks become a running score per model, and the loser’s file stays on disk.

Versus for whole conversations

A lighthouse on a rocky point at first light

The stage during a Versus: the same lighthouse prompt made by google/gemini-2.5-flash-image on the left and openai/gpt-image-1 on the right, each with a “This one wins” button underneath.
03

Edit what came back.

Paint a mask and redo part of a picture, edit it by prompt, cut the background out, upscale it 2× or 4×, or extend it to one of six aspects. Each edit writes a new file named after its source, so the original is kept. Cutting the background out is not a generation: the cut runs on your machine, and the result is checked for real transparency before it says so. Vector results open as layers you can recolour or delete.

The Edit image window over a harbor picture, on Inpaint, with the other modes beside it: Edit by prompt, Remove BG, Upscale and Outpaint. A brush slider, the line “Paint over the part you want changed. Everything else stays as it is”, a box for what goes in the painted area, and Apply edit. The footer says edits are best-effort and the original is kept.
04

A camera you choose, written into the prompt.

Six camera bodies, eleven lenses, six focal lengths and three apertures for any image or clip, and 36 camera moves for video. They are added to your words when you press Generate, so the journal holds the exact prompt the model was sent, not a hidden suffix.

The Camera setup window. It says the choices are added to your prompt when you generate and your own words stay as you wrote them. Six camera bodies, each with a one-line description, Full-Frame Cine Digital ticked, and the lens list beginning below.
05

A file, a price and a journal row.

The Studio runs the same media.* commands as an agent asked in chat, so both kinds of generation look the same in the journal. The file lands in the project’s media/ folder. The cost shown is what the provider reported. Delete is the generation’s own undo, and it can only remove media files.

The Studio and the chat call one command

make a hero image for the launch post

The Studio’s stage showing a generated harbor scene, the prompt that made it, the model google/gemini-2.5-flash-image, the size 1600×900, the file path in the project’s media folder and a cost of four cents, with Edit, Animate, Use as reference, Remix, Save prompt and Download under it.
06

Talk to the composer. The audio is thrown away.

The mic on the composer records, then a transcription model you pick turns it into text in the box. The chat model never hears the recording, and the recording is never written to disk: the transcript is the only thing kept.

The composer with its mic button beside Send, and below it Settings, Voice, Dictation: the mic records, then this model turns it into words, separate from the chat model, which never hears the recording. Picked by ranking: openai/whisper-large-v3, $0.006 per minute, measured.
What it will not do

A video job that runs past 20 minutes is reported as timed out and is never sent again, because a second submit is a second bill. How undo is bounded.

At a glance

The Studio, in numbers.

Images per request
Up to 4, guided by up to 8 reference pictures.vb-llm/src/media.rs, vb-cmd-media/src/generate.rs
Video
1 to 60 seconds as the model allows; up to 16 references, or start and end frames, not both.vb-cmd-media/src/generate.rs
Video wait
Polled every 10 seconds for up to 20 minutes. A timeout is final.vb-cmd-media/src/generate.rs
Prompt length
1,500 characters in the panel, 4,000 at the command, 8,000 for speech.ui/src/Studio.tsx, vb-cmd-media/src/generate.rs
References
Read from your disk and sent inline as base64, up to 20 MB each.vb-cmd-media/src/store.rs
Where files land
<project>/media/<prompt-slug>.<ext>, the name cut at 48 characters; with no project open, the library.vb-cmd-media/src/store.rs
Cost
The provider’s own reported cost, per image or video generation, never estimated from the catalog.vb-llm/src/media.rs
Presets
37 styles, 6 camera bodies, 11 lenses, 6 focal lengths, 3 apertures, 36 camera moves.ui/src/studio/
Music controls
8 genres, 7 moods, 4 vocal styles, lengths of 15, 30 or 60 seconds, BPM and an Avoid line.ui/src/studio/music.ts
Saved prompts
Up to 100, kept per tab: image, video, music, speech.ui/src/studio/vault.ts
Dictation
Audio up to 32 MB in wav, mp3, flac, m4a, ogg, webm or aac, transcribed and never saved.vb-cmd-media/src/listen.rs
Delete
Up to 10 files per discard, and only media files under a project’s media/ folder.vb-cmd-media/src/gallery.rs

The details,
for the careful.

What can the VectorBrain Studio make?

Images, video clips, music and speech, from whichever models your OpenRouter key reaches that produce them. Dictation goes the other way and turns your voice into text in the composer.

Which models can I use?

Any model in the OpenRouter catalog that outputs the thing you are making. The Studio lists them with their published prices, and a model added to OpenRouter today appears after the next catalog sync. How models and keys work

Why do the controls change when I pick another model?

The controls are read from the model’s published capabilities. A setting the new model cannot produce snaps back to Auto instead of being sent and ignored. A model that publishes no capabilities gets permissive generic controls.

Is a Studio generation different from asking in chat?

No. Both call the same media commands with the same input, both write a journal row, and undo works the same way on both. A connected companion such as Claude Code can use the same commands if you allow it. The Studio from your terminal

Can I undo or delete a generation?

Yes. Delete runs the generation’s own undo, which removes its files from disk permanently after you confirm. It can only delete media files inside a project’s media folder. It does not refund what the provider charged.

How do reference images reach the model without a server?

A desktop app has no public URL to hand a provider, so references are read from your disk and sent inline as base64 data, up to 20 MB each. A provider that only accepts a hosted URL cannot use them.

What does a generation cost?

What your provider charges your key; VectorBrain adds nothing. Image and video generations show the cost the provider reported. Music and speech do not record a cost yet.

Where does dictation send my voice?

To the transcription model you picked, through your key, and nowhere else. The recording is never written to disk and the chat model never hears it. Its price per minute is measured, from a three-second test clip or from your own transcriptions, because OpenRouter does not state the unit.

Can I export documents from here?

Documents have their own page. They export to PDF, DOCX and PPTX. Documents and export

Your key, your folder.
Every picture has a receipt.

What the Studio makes is a file in your project, with the prompt, the model and the price written beside it.