FLUX.2 Klein 4B through Core

The flux2-klein-4b model definition selects distilled FLUX.2 Klein 4B for text-to-image generation through the shared local image runtime. It uses the same retained native context, progress, cancellation and complete PNG result contract as other selected image models. Applications own prompts, image associations and saved results.

Model selection

Select this runtime record under native.localAI in arcane-app.json, or the corresponding package projection's localAI record:

{
    "runtimes": [
        {
            "id": "stable-diffusion.cpp",
            "version": "master-929-3f8527a",
            "backend": "auto",
            "models": ["flux2-klein-4b"]
        }
    ]
}

The definition is a single FLUX2_KLEIN_4B_MODEL record in src/local-ai/image-models/flux.mjs. The built-in model ID supplies that record through normal runtime configuration; the shared service exposes its resource metadata through image.status and images.inspect().

Backend selection belongs to the shared runtime. This definition adds no CPU offload setting or device override. auto uses the available runtime's accelerated backend with CPU fallback; actual support depends on the installed runtime distribution. See the shared guide for platform distribution availability. Selecting the model does not establish that it has loaded or run on a GPU, CPU or NPU.

Required model resources

Prepare all three resources together. They are complementary parts of one model pipeline, rather than three independent image generators.

The selected upstream records identify these resources as Apache-2.0. The engine supplies its embedded Qwen tokenizer. This definition needs no separate tokenizer download, chat server or LoRA. The full encoder in the selected VAE also retains the model's image-encoding capability; a decoder-only file is a different resource.

The existing DBOPFS model store owns authoritative model files and resumable downloads. The shared model-assets service prepares complete native working files. Runtime installation does not download another copy of these models. Resource roles map to the exact filenames returned by that preparation.

The following extends the shared generation guide. It uses an existing images accessor, DBOPFS-backed modelStore, application-selected workingDirectory and request signal:

import {createBrowserModelSource} from 'arcane-os/ai/browser-wasm';
import {prepareCoreModelAssets} from 'arcane-os/ai/core-model-assets';

const selected = (await images.inspect({signal})).models.find(
    function selectedKlein(model) {
        return model.id === 'flux2-klein-4b';
    }
);
const resources = Object.entries(selected.resources);
const source = createBrowserModelSource({
    id: selected.id,
    files: resources.map(function modelFile([role, resource]) {
        return {name: resource.filename, url: resource.url};
    })
});
const stored = await modelStore.ensure(source, {signal});
const projection = await prepareCoreModelAssets({
    workingDirectory,
    members: source.files.map(function workingMember(member, index) {
        return {path: member.name, file: stored.files[index]};
    }),
    signal
});

try {
    await images.load({
        model: selected.id,
        assetProjectionId: projection.id,
        resourcePaths: Object.fromEntries(
            resources.map(function nativeRole([role, resource]) {
                return [role, resource.filename];
            })
        ),
        signal
    });
} finally {
    await projection.release();
}

const result = await images.generate({
    model: selected.id,
    prompt: 'A dignified raccoon astronaut watering a moon cactus in a teacup.',
    parameters: {width: 768, height: 768},
    signal
});
// Display result.images using the shared guide's Blob preview example.

Change the prompt to request another scene and the dimensions to select its canvas. Successful loading retains the working projection inside the native model owner; releasing the caller's preparation handle leaves that native retain intact. Repeated generation reuses the loaded context. Explicitly unload it when the application no longer needs it, following the shared lifecycle contract.

Defaults and model behavior

The definition supplies four sampling steps, the Euler sampler and sample_params.guidance.txt_cfg: 1. Omitted native options retain their native initializer defaults, including automatic FLUX.2 scheduler selection. These settings belong to the distilled 4B checkpoint; the separate Klein base and 9B variants have their own model requirements.

The native default canvas is 512 by 512. Klein's native preparation aligns dimensions upward to multiples of 16; use the returned image dimensions when displaying or saving results. The example already selects aligned dimensions. At the default CFG value of 1, the native text-to-image path does not apply negative-prompt conditioning.

The SDK forwards the caller's complete prompt to the native API. The engine then applies its prompt-weighting syntax and Qwen conditioning format. Parentheses and brackets can affect weighting. The selected native API has no public switch for literal text conditioning, so exact SDK transport does not imply literal interpretation by the model. The shared guide describes inputs that cannot pass unchanged through its native C-string boundary.

This definition advertises txt2img. The upstream Klein model also supports reference-image editing, while the SDK's separate whole-image img2img operation uses a different input path. Reference editing becomes an SDK capability when the shared runtime implements its reference-image dispatch; selecting this model alone does not add that operation.

These settings and resource roles follow the tagged FLUX.2 instructions, the official Klein model, the official small decoder and the official Qwen model files.