An API-format harness, not a canvas workflow — read this first

This JSON will not drag onto the ComfyUI canvas. That is deliberate, not an oversight. It is a reproducibility harness driven by a script, so every run is byte-checkable against the last one. If you want a graph to click around in, this is not that.

The claim, stated precisely

Re-running this harness unchanged on the same machine reproduces the reference image with a mean absolute difference of 0.0000 — bit-identical. Not "close". Identical.

Commercial-safe, verified against primary sources

The harness loads three files and all three are Apache 2.0:

• flux-2-klein-base-4b.safetensors — Black Forest Labs

• flux2-vae.safetensors — Black Forest Labs

• qwen_3_4b.safetensors (Qwen3-4B) — Alibaba / Qwen

The 4B/9B split matters. FLUX.2 klein 4B Base is Apache 2.0. 9B Base is NOT — it ships non-commercial. Swap in 9B and the commercial-safety of this stack no longer holds. Verify yourself rather than taking my word: the HuggingFace model cards and LICENSE.md in the BFL flux2 repo.

Apache 2.0 is permissive, not public domain — redistribution requires the licence text and attribution. Credit BFL and Qwen accordingly.

Reproducibility traps — read before reporting a mismatch

1. Use the full-precision file, not fp8. fp8 produces different pixels. Same seed, same prompt, different numbers.

2. Note the nested path. The UNET loads from a split_files/diffusion_models/ subfolder. If your install is flat, nest the file or edit unet_name.

3. Do not change steps, cfg, sampler, or resolution. 30 / 5.0 / euler / 1024x1280. Every one is part of the lock.

The method, in one sentence

Hold the seed and every other token identical, and swap only the pose phrase. The lock is the whole prompt, not just the seed.

Known limits — named here so you don't have to find them

This ships with documented defects. A clean demo that hides its failures is worth less than a defective one that explains them.

1. The pose phrase leaks into the face. A high-energy pose rewrote the character's face with no facial term changed in the prompt. The more energetic the pose, the more prompt weight the face needs. The failure case is included as a labelled figure rather than quietly deleted.

2. Naming a region without bounding it invites drift. The prompt says "bare cream fur chest and belly" but never says where that region ends, and the cream area terminates at different heights across poses. Bounded, explained, and deliberately unfixed — fixing it forces a full re-lock.

3. Determinism means you cannot re-roll. Same seed plus same prompt equals the same pixels, forever. There is no "try again". A genuine trade-off of the method, not a bug.

Licensing — three separate layers

• Model stack — Apache 2.0, commercial use permitted (see above)

• The harness, script, and documentation — MIT. Commercial use explicitly granted. Build a paid character kit with this; that is the point.

• The demo images — CC BY-NC 4.0, attribution ytcreator432327.

Why the images differ: the red panda is a brand asset. The method is given away completely and commercially. The character is not. Build a commercial character kit with this harness — you just can't ship my red panda.

On AI images and copyright: the status is unsettled in most jurisdictions. This grants whatever rights exist rather than asserting exclusive ownership I may not hold.

Full write-up and source

Method walkthrough, the v1-v3 case study, and the harness itself: https://github.com/zoomg333/pickaxe-consistency-kit

Same person, two handles: Civitai ytcreator432327 and GitHub zoomg333. Stated so the repo and this listing verify against each other in both directions — a release that asks you to trust its measurements should not have an unverifiable author.

If this is useful, follow or leave a comment. I'm gauging whether there's demand for a fuller commercial-safe character kit before building one. Questions about commercial use, variations, or a paid version are exactly the signal I'm looking for — ask them here.