Fast Image Generation

The Problem

Standard on-device image generation takes 15–30 seconds. The highest-quality open-weight models (FLUX, SD3-Large) don't even fit in memory on most consumer hardware; on the rare machines where they fit, a single image can still take minutes to generate. Cloud APIs side-step the memory ceiling but add network latency and recurring per-image costs. None of that works for real-time game workflows. And cloud generators raise an additional artist-consent problem: they train on uncontracted artwork, which poisons the value chain.

Edit the prompt, get a new image: 512×512 in about 1 s with one step on an M1 Max.

Performance

Atelico IG, our proprietary on-device image-gen model, generates a 512×512 image in about a second on an M1 Max and well under half a second on an RTX 3090, at output quality close to FLUX.2 [pro]. For context, the chart also shows two cloud baselines from Replicate measured end-to-end (queue + inference + network): FLUX schnell at ~0.8 s and FLUX.2 [pro] at ~10.8 s, both at 1024×1024.

Atelico IG fits in the unified memory budget of every consumer device we ship to, including phones; the cloud-served FLUX.2 [pro] is a 32B-parameter model that does not.

Six Atelico IG renders with the same seed as the prompt grows: a dog; a dog detective; wearing a hat; and a bow tie; with glasses in the rain at night; with neon lights.

Game Developer Use Cases

Runtime asset generation: procedural item icons, character portraits, loading screens
In-editor tooling: rapid concept art iteration, texture generation
Player-facing: user-created creatures (like Bobium Brawlers), custom avatars
No cloud costs: runs entirely on the player's device

Adapter Support

Hot-swap style adapters at runtime to match your game's visual identity. Currently powered by LoRA, adapters let you train on your concept art and load the results at runtime with zero downtime.

let image = engine.images().generate(
    "a crystal sword with blue glow",
    Some("fantasy-style-lora")
).await?;

A hound detective in a fedora and trench coat on a rainy night street, in black-and-white film-noir style. — Film noir

A hound detective in a deerstalker with a pipe at a desk of maps, as a nineteenth-century oil painting. — 1800s oil painting

A golden retriever detective with a magnifying glass on foggy London cobblestones, in a polished cartoon style. — Cartoon

A beagle detective in a deerstalker and trench coat holding a magnifying glass, in a flat 2D chibi style. — Flat 2D chibi

Adapters Trained on Artist-Approved Images

We partner with artists and train LoRA adapters on their approved imagery to capture their style. No scraped artwork, no uncontracted datasets — royalties flow back to the artists whose work shapes each adapter, so the value chain stays intact instead of leaking out to a third-party API.

Mobile Support

Runs natively on iPhone 13 and newer. The compact architecture of Atelico IG is purpose-built for mobile constraints, delivering high-quality results without exhausting device memory.

Generate images in ~1 second