FLUX.2 [klein] 9B Inference¶

★★★★★ Intermediate

FLUX.2 [klein] is a model family. FLUX.2 [klein] 9B is a named distilled release, but it is still not a complete inference configuration without its checkpoint revision and runtime assets. The official FLUX.2 repository distinguishes 9B, 9B KV, and 9B Base variants; its guidance positions distilled variants for interactive workflows and base variants for fine-tuning and maximum flexibility. Treat the named checkpoint and its published runtime assets as one contract.

Identify the exact contract¶

Record before an inference or editing run:

model and checkpoint revision, including distilled/base or KV variant;
pipeline, scheduler, text encoder, decoder/VAE, adapter loader, and runtime;
prompt or image-reference fixture, seed, dimensions, and output format;
all adapters and their source/base-model compatibility; and
the applicable weights, API, and derivative-use terms.

The official project lists FLUX.2 [klein] 9B under a non-commercial license. That does not authorize a product workflow by itself: verify the current terms for the selected weights and separately verify any hosted-service or third-party component terms.

Text-to-image and editing are separate tests¶

For text-to-image, use a small, versioned prompt suite that includes composition, material, typography, and negative fixtures. For image editing, retain the source asset and declare the requested changes and protected properties before generation.

The FLUX.2 image-editing documentation describes text-directed single- and multi-reference editing. It does not mean that every local wrapper, checkpoint, or adapter preserves all details. Check the exact route with source/edited comparisons, not with a general claim about model capability.

Tune by measurement, not folklore¶

Do not copy a universal step count, guidance value, sampler, denoise strength, VRAM figure, or LoRA-weight range from another workflow. Instead:

run the published baseline for the chosen checkpoint;
vary one supported setting at a time on the same fixtures;
measure cold and warm latency, peak memory, reproducibility, and visual acceptance;
evaluate source preservation separately for an edit workflow; and
publish the selected version tuple and known limits with the job.

Quantization, compilation, and memory offloading are runtime trade-offs. See diffusion inference acceleration for the receipt required before they are combined with an image workflow.

Adapter compatibility¶

An adapter must name the exact base checkpoint and loader it was trained against. Similar parameter counts or a shared “FLUX” label are not evidence of compatibility. Test an adapter alone before testing a combination, keep its scale and load order in the run receipt, and reject an output if its protected content drifts.

For training a new adapter, use the current model-specific guidance such as the FLUX.2 [klein] training documentation, then validate the resulting artifact under the same inference pipeline that will serve it.

Release record¶

Store the model manifest, prompt/reference fixture, output hash, acceptance review, license decision, and any output-marking metadata with each approved run. This makes a later runtime or model upgrade observable instead of silently changing image behavior.