FLUX.2 [klein] 9B Identity LoRA: Disentanglement Contract¶
An identity LoRA learns an adapter for a particular base-model release; it does not isolate a person from every surrounding attribute by default. “Disentanglement” is a preservation claim that needs evidence for the named data, adapter, runtime, and prompt/edit workflow.
The official FLUX.2 [klein] training documentation documents LoRA fine-tuning for the family, including character-consistency use cases. It also distinguishes the 9B base model and its terms from other variants. That supports an experiment on the exact compatible release, not a claim that block choices, rank reduction, adapter arithmetic, or another family's recipe transfers unchanged.
Bind the adapter and its authority¶
For each run, retain:
- official base-model/checkpoint identifier, artifact digest, license and access terms, model runtime, and adapter serialization/loading path;
- authorized reference assets, consent or other usage authority, purpose, retention/deletion rules, captions, trigger vocabulary, and any excluded attributes or regions;
- training split, source/derivative grouping, preprocessing, crop/orientation, caption policy, augmentation policy, code/config revision, and seeds;
- proposed adapter file, merge/strength controls, output digest, and the exact workflow used for evaluation; and
- reviewer decision with identity-fit evidence, non-target preservation evidence, failure examples, and a permitted-use conclusion.
A 9B adapter must be paired with the model release and runtime it was trained for. Do not infer compatibility from a similar FLUX variant, a file extension, or a community UI label.
Measure separation rather than assume it¶
Define which requested identity properties may transfer and which non-target properties must remain controllable: pose, expression, age presentation, clothing, background, lighting, style, body, and protected attributes. Evaluate the requested transfer separately from preservation on held-out, source-disjoint prompts and references. Keep authorized reviewer corrections and negative cases, especially where an adapter repeats a background, changes a protected region, or creates an unintended likeness.
Block targeting, rank selection, orthogonality losses, subtraction, merging, and SVD truncation are model- and implementation-specific experiments. They need a declared hypothesis, pinned implementation, comparison baseline, and the same identity/preservation review; none is a universal cleanup step.
Failure boundary¶
If the base release or terms are unknown, reference authority is incomplete, the adapter/runtime format differs, source groups leak into evaluation, or a reviewer cannot distinguish requested transfer from non-target change, keep the adapter out of use. Do not silently load it on another variant, remove evidence, or label a generated likeness as verified identity.