PILOT / FUNCTIONAL TESTING
Blackwell Diffusers Lab
A working notebook for FLUX.2 Klein image-to-image LoRA caption dropout. Inspect each execution, its limits, and the evidence behind it.
Read the record.
Keep the failed runs.
Make it reproducible.
Bounded functional tests of a real pretrained model on programmatic image fixtures. Loss and memory are diagnostics. They do not establish training quality, steady-state performance, or a speedup.
Read the method ↓Checking the evidence package…
METHOD & LIMITS
What this evidence can support
A functional pilot
The question is whether the bounded training and adapter checks execute under the recorded configuration. Programmatic image fixtures exercise the code path; they do not evaluate image quality or establish a useful fine-tune.
Only observed loss points are drawn. Missing measurements remain missing, with an explanation. Memory uses a zero baseline and the selected run's own scale. Time and memory are descriptive diagnostics; no cross-run ranking, percentage improvement, or performance comparison is calculated.
Trace a value to its source
Open a numeric value to inspect its raw JSON evidence. The metric table includes the exact traversal key. Each run also exposes the validated record, original artifact hashes, exact model/source revisions, configuration and dependency versions.
A passing data audit checks the supplied records. It is not a release approval, a reproduction by another machine, or maintainer acceptance. Independently review the commands, licenses and claims before redistributing this package.
Why there is no model or benchmark dataset here
The artifact is a reproduction and test-harness notebook. Model checkpoints and LoRA adapter weights are deliberately withheld. A small functional pilot does not justify a benchmark dataset or a general-purpose fine-tune claim. Public evidence downloads contain text records and reproduction code, not model weights.
REPRODUCTION
Take the experiment apart.
Use the per-run configuration and pinned revisions. Restore each downloaded file to its recorded relative path in a clean workspace. Review the scripts before running them; GPU execution and model downloads happen on your machine.
Upstream coordination
No PR opened. Maintainer acknowledgment and human-authored wording are pending.
This notebook is independent work. The issue discussion and earlier proposal by chinmayirm are prior art, not an endorsement of this implementation.