Just Say What You Want: How the XBRUSH Workspace AI Agent Handles Image, Video, and Audio
The XBRUSH Workspace has a lot of functions. Image generation, editing, inpaint, outpaint, background removal, upscaling, video generation, audio generation, style training. The problem is knowing which one your task lives under. If you have to learn the tool before you can produce anything, most people never start.
The Workspace AI agent inverts that order. Picking the tool is the agent's job. You describe the result.
Why "you don't have to know the product to use it" actually holds
The key is that the agent knows every XBRUSH function at the API level. Unlike a person hunting through menus, the agent treats each function as something it can call directly.
So one line of plain language triggers four things at once.
- Read the intent — figure out what is actually being asked for
- Pick the tool — generate or edit, image or video or audio
- Pick the engine — the AI model that fits the job
- Write the prompt and execute — rewrite it in the form the model understands, then call it
Steps 2 and 3 are exactly what you no longer have to learn. Which model is strong at which task, how editing engines differ from generation engines — that judgment sits on the agent's side.
Two ways to work, on one screen
The Workspace keeps two control modes side by side.
- Chat — describe the outcome and the agent decides the method
- Left panel — set Image/Video/Audio tabs, generate/edit/outpaint, AI engine, image size, and image count yourself
It is not an either/or. You can let the agent draft, then change only the one setting that bothers you in the left panel. Automation and control are not a trade-off here.
There are two ways to begin, too: type the result you want, or drag a file in. A dropped file becomes source material immediately.
A real job: "make me a 9-frame sprite sheet"
The documentation includes a session that shows the whole flow. The user uploaded one cat archer character image and wrote — in Korean, which is worth noting, because the prompt the agent ended up sending to the model was in English:
I want to turn this cat archer's shooting scene into a 2D sprite for a game. Make me a 9-frame 2D sprite.
Break the agent's response down and you get:
- It states the plan first — "I'll create a 3x3 nine-frame archery sprite sheet"
- It announces engine and cost up front — gpt-image-2-edit, low quality, roughly 2.4 credits
- It writes the English prompt itself — nine frames of drawing, aiming, releasing, and follow-through, keeping the same character design, consistent size, side view, and a flat clean background, all generated automatically
- It reports the result and the actual charge — "Your 9-frame (3x3) archery sprite sheet is done (2.4C used)"
- It offers the next step — "Want me to split it into individual frame files?" One more line back ("yes, split it") and the agent cuts the sheet into 9 cells without any of it being re-explained
Two details are worth pausing on.
First, the user never said "edit mode." Yet the agent chose editing, not generation, because the uploaded character had to survive the job. Generate from scratch and the cat archer becomes a different character.
Second, the cost is announced before the credits are spent. You are not reading a bill after the fact.
What the agent can take on
The sprite sheet is one example. Most Workspace functions can be driven by conversation.
| Area | Example instruction |
|---|---|
| Image generation | "Give me 4 product concept shots" |
| Image editing | "Change the background to a beach", "Remove that person" |
| Background removal · outpaint | "Cut this out cleanly down to the hair", "Extend it wider" |
| Upscaling | "Bump the resolution for print" |
| Video generation | "Add motion to this photo" |
| Audio · music | "Make an upbeat background track" |
| Sprite sheets | "Turn this into a 9-frame 2D sprite" |
| Style training (LoRA) | "Build a style out of this look" |
The practical difference is not a single image. It is that image to video, video to audio all happen inside the same conversation.
Keep talking and the work keeps going
Once a result comes back, the next instruction picks up right there.
Make the background brighter
Now turn this into a video
You never re-explain from scratch, because the earlier result stays in context.
Outputs accumulate inside a task, and tasks live in folders. Split them by project, client, or campaign and they stay findable later.
Teams work the same way. A teammate opens the same task, adds instructions, and results stack on top — with a conversation history showing who ran what.
From one-line jobs to full deliverable sets
If you are not sure what to say first, there are two prepared starting points.
Quick Start — 20 single-instruction jobs
Edit by Chat, Remove Background, Studio Product Shot, Create an Image, Photo to Video, Upscale, and more. Each card carries an example phrase you can copy verbatim. A rotating weekly style pick changes every week.
Workflow Recipes — 21 deliverable sets
These bundle several steps into one complete set of outputs. Every card shows the step count, the number of outputs, and a credit range.
The Space Promo Kit, for instance, runs seven steps.
- 3 tone directions · 2 hero shot retouches · 2 virtual styling options
- 1 info card · 5 Instagram carousel images · 1 story notice · 1 wide cover
- Roughly 15 images · roughly 36–311 credits
Starting materials: one to three photos of the space, its name, address and hours, and a one-line description. Once the task is created, the agent asks for each material in order. You supply the inputs; the agent manages the sequence.
The same functions in code — the XBRUSH Public API
Saying the agent works at the API level is not a metaphor. The same functions are exposed publicly through the XBRUSH Public API.
| Group | Endpoints |
|---|---|
| Images | POST /images/generate · /images/edit · /images/upscale · /images/background-remover |
| Videos | POST /videos/generate · /videos/upscale |
| Requests | GET /requests/{requestId} plus move and delete |
| Sessions · folders · outputs | /sessions · /folders · /outputs |
| Utilities | GET /models · GET /balance |
What to know before you build on it:
- Everything is asynchronous — a generation request is accepted with 202 and a
requestId. Poll untilstatusiscompleted, then use the output URLs. Images never come back inline. - Sessions are mandatory — every generation request needs a
sessionId. A session is the same thing as a task in the web Workspace, which is why API results show up there too. - Shared credits — team credits are charged on submission and automatically refunded in proportion to whatever fails.
- Rate limit — 60 requests per minute per key, applied to generation submissions only. Reads are unrestricted.
- Auth — the
X-API-Keyheader. Keys are team assets issued by an owner or admin in team settings, and belong server-side only.
In short, chatting and calling code hit the same engines, the same credits, and the same workspace. Whatever flow you validate by hand transfers straight into automation.
Four ways to instruct it well
1. Describe the result, not the operation
"Extend it to 16:9 with outpaint" is worse than "vertical, for an Instagram story." Name where it will be used and the dimensions follow.
2. Drop the material in instead of describing it
Uploading a file beats describing an image in words. For anything where identity matters — a character, a product — always upload.
3. Don't ask for everything at once
Pull one frame first, then say "the rest in this tone." Order all 15 up front and a wrong direction costs you all 15.
4. Read the cost preview before you approve
The agent tells you the engine and the estimated credits before running. If the quality tier looks too low, that is the moment to ask for a higher one.
Frequently Asked Questions
Can I use this if I'm bad at writing prompts?
Yes. Describe the result in plain language and the agent writes the prompt the model needs. In the sprite sheet example the user wrote a single sentence, and the agent produced the English prompt containing all the specifics — 3x3 grid, consistent character design, side view, flat clean background.
Can I see the credit cost before it runs?
Yes. The agent states which engine and quality it will use and the estimated credit cost before executing, then reports the actual amount charged when it finishes. Workflow Recipes display a credit range on the card itself — the Space Promo Kit, for example, shows roughly 36–311 credits.
What if I don't like the engine the agent picked?
Use the manual panel on the left to set the AI engine, image size, image count, and the generate/edit/outpaint mode yourself. Chat and the manual panel sit on the same screen, so you can mix them — let the agent draft, then change only the setting you disagree with.
Can my team work on the same task together?
Yes. Teammates can open the same task and keep giving instructions, with results accumulating in place. The conversation history records who ran which step, so the work is traceable. Outputs are organized into tasks, and tasks into folders.
Are the same functions available through the API?
Yes. The XBRUSH Public API exposes image generation, editing, upscaling, and background removal, plus video generation and upscaling. All requests are asynchronous — submissions return 202 and you poll the requestId — and generation requests require a sessionId. It draws on the same team credits, and outputs appear in the web Workspace as well.
Tools used
- XBRUSH Workspace AI agent — plain-language instruction → tool and engine selection → prompt writing → execution
- Manual panel — Image/Video/Audio tabs, generate/edit/outpaint, engine, size, and count set by hand
- 20 Quick Start templates — single-instruction jobs
- 21 Workflow Recipes — multi-step packages producing a full deliverable set
- Tasks · folders · team collaboration — accumulated results, conversation history, per-project organization
- XBRUSH Public API — the same functions from code, sharing team credits and the Workspace