Noise in.
Picture out.

Your sentence becomes a compact representation of meaning. The model starts from pure noise and removes it in steps, steered by that meaning, until an image is left behind.

A luminous crystal core visualising Paintune's latent diffusion technology.

Architecture

Diffusion, run in a compressed space

Denoising happens in a latent space rather than on raw pixels. That is what makes high resolution affordable to generate instead of theoretically possible.

Latent diffusion with a transformer text encoder
Trained and evaluated in house, not wrapped
Deterministic seeds throughout
Tuned against hands, in-image text and style drift

Property	Value
Architecture	Latent diffusion, transformer backbone
Conditioning	Transformer text encoder
Sampler	Flow-matching scheduler
Output	Up to 2048 px · PNG, JPEG, WebP
Reproducibility	Settings on every response
Provenance	C2PA credentials in the file

Pipeline

Five stages, all of them logged

Every generation can be explained after the fact, not just produced.

Prompt processing

Your text is parsed, checked against policy, and encoded into a meaning vector.

Latent init

A noise tensor is seeded. This seed is what makes the result reproducible.

Denoising

Noise is removed across the requested steps, guided by the encoding.

Decode & upscale

The latent becomes pixels and is taken to delivery resolution.

Sign & deliver

Credentials are attached and the file returns with its settings.

The schedule

Structure first, detail last

A lighthouse, almost entirely obscured by noise. — One frame across the schedule. Each step predicts the noise present and subtracts a fraction of it, so structure appears before detail does. *Illustration — the degradation here is applied for the diagram, not a capture of the intermediate tensors.*

The same lighthouse, still heavily degraded but with shapes emerging. — One frame across the schedule. Each step predicts the noise present and subtracts a fraction of it, so structure appears before detail does. *Illustration — the degradation here is applied for the diagram, not a capture of the intermediate tensors.*

Latent space visualised as converging magenta and blue particles.

Infrastructure

Bursty by nature, so built to scale

Nothing for an hour, then forty images in two minutes. The platform is built for that shape rather than for holding idle capacity.

GPU-accelerated inference
Autoscaling capacity, queued batch execution
Per-request logging of settings and policy checks

Safety

What we do about misuse

Filtering both ways

Requests are checked before generation and results before delivery. Prohibited categories are listed in the acceptable use policy.

Credentials in the file

C2PA metadata records that the image was generated, by which model, and when. It travels with the file rather than living in a database.

No likeness training

We do not train on datasets built to reproduce a specific living person, and we refuse prompts whose purpose is to impersonate one.

Read the API reference

Same model, reachable from your own application in three endpoints.

Start creating Contact us

Noise in.Picture out.