FLUX.2 Klein in ComfyUI: A Reproducible Setup Record

By Dominic Marrocco, 42 UK Research. Technical claims below are bounded to the linked primary sources.

Irish reproducibility edition · evidence review: 5 August 2026

Turn a local run into reproducible evidence

This companion withdraws the old two-gigabyte hardware assertion. The 4B base publisher card says the model fits in about 13 GB of VRAM, while the 9B publisher card says its different, four-step distilled release fits in about 29 GB and uses a non-commercial licence. A reproducible record must name which one was used.

Minimum run record

1. Record the exact Black Forest Labs repository and model revision.
2. Record the ComfyUI release or commit obtained through the official project.
3. Record the operating system, accelerator, driver, Python and PyTorch versions, using the official requirements as the compatibility reference.
4. Archive the workflow JSON, input assets, seed, dimensions, batch size and launch flags.
5. Define whether reported memory is allocated, reserved or device-level peak memory and name the measurement tool.
6. Keep failed runs and error messages; removing them creates a misleading success-only record.

When the workflow does not fit

ComfyUI’s official troubleshooting page lists memory-pressure controls including --lowvram, lower resolution or batch size, disabled previews and reserved-VRAM adjustment. Record the before-and-after configuration. These mitigations are variables in the test, not evidence for an unmeasured 2 GB claim.

Publish through the evidence hubs

Attach raw memory observations to the benchmark methodology, register a reproducible setup in the installation matrix, and publish workflow assets with a declared trust level in the workflow library.