One model. Four surfaces. Built to scale on demand.
Studio to create in, an editor to finish in, a brand layer for teams and an API for engineers — all on the same pipeline, with the same safety layer and the same privacy guarantee.
The four surfaces
Studio
CreateThe text-to-image workspace. Write a prompt, pick a style family, get four studio-quality variations, and iterate in real time while previews stream. Seeds and negative prompts are one click away when a frame is nearly right.
Smart Editor
RefineFix the frame you have instead of rolling the dice again. Inpaint a region, extend the canvas past its original edges, swap a background, erase an object — the pixels around your edit stay untouched.
Brand Engine
BrandFine-tune Metabob on your own products, palette and lock-ups, then generate campaign assets that stay on identity across a whole set. Adapters train from as few as a dozen references and stay isolated to your workspace — never merged into the base model.
Developer API
ShipGenerative imaging as an API call inside your own product. Predictable REST semantics, streaming responses, idempotent job submission and signed webhooks, with SDKs for JavaScript, Python and Go.
| Endpoint | Does | Returns |
|---|---|---|
POST /v1/images | Generate from a prompt, style family and ratio | Job id, then streamed previews |
POST /v1/edits | Inpaint, outpaint or erase within an existing frame | Job id, then the edited frame |
POST /v1/variations | Re-roll around a chosen seed | Job id, then n variations |
POST /v1/upscale | Detail-recovery pass to 4K | Job id, then the upscaled asset |
GET /v1/jobs/:id | Poll a job when you would rather not hold a stream | Status and asset URLs |
POST /v1/finetunes | Train a brand adapter from a reference set | Adapter id and training status |
This is the endpoint contract we publish for integrators, so teams can design against it before they write a line of code. Request access and we will set you up with keys.
Architecture
A multi-model pipeline: we route each prompt to the right model, refine composition and detail, then stream the result back. Inference runs on a pooled fleet of cloud GPUs behind an autoscaling scheduler, renders land in object storage and serve from a global edge cache — so capacity follows demand instead of running ahead of it.
Latent diffusion core
A latent diffusion backbone producing both photoreal and stylised output, with a refinement pass that recovers detail at 4K. Nine trained style families, stackable on one prompt.
Multimodal text encoders
Hybrid encoders capture style cues and compositional intent, so a longer prompt adds control rather than noise.
Autoscaling GPU fleet
Inference runs in containers scheduled across a pooled GPU fleet, with a warm pool sized to live traffic so queue depth stays flat at peak instead of spiking with it.
Streaming delivery
Render jobs move through a durable queue, land in object storage and serve from a global edge cache, so a preview reaches the browser while the full-resolution frame is still finishing.
Safety and IP filtering
Layered moderation runs on every generation — NSFW, trademark and likeness checks — before an image is ever returned to a user.
Brand fine-tuning
LoRA adapters train per customer on a small reference set, kept isolated per tenant and never mixed into the base model.
What we do with your work: nothing
The commercial case for a generative tool collapses if customers cannot trust it with a brief. So the guarantees are structural, not policy footnotes.
Your prompts, uploads and outputs are private to your account and are never used to train or evaluate our models. There is no opt-out to find, because there is no opt-in.
A brand adapter trained on your references belongs to your workspace. It is never merged into the base model and never serves another customer's prompt.
NSFW, trademark and likeness screening runs inline before an image is returned — not as a nightly sweep after the fact.