The FLUX LoRA Standard Already Picked Itself
When Replicate builds their commercial product on your repo, the debate is over.
ai-toolkit stayed the standard FLUX LoRA trainer. Replicate kept building on it. The ecosystem vote held.
For a while the answer to "how do I train a LoRA" was kohya. kohya-ss/sd-scripts was just the thing — messy, sprawling, held together with enough shell scripts and community documentation that it worked anyway. If you were fine-tuning anything in the Stable Diffusion universe you were almost certainly in there somewhere.
That's over now.
ai-toolkit by ostris is the de facto FLUX LoRA trainer, and the tell is Replicate — they built their LoRA product on top of it. When a company that charges money for a thing picks your library as the foundation, that's not an endorsement, that's a verdict. The ecosystem vote happened without a meeting.
The license matters too. MIT, which means actually permissive — not "permissive except for the part where you read the README and find out it isn't." kohya's licensing situation has historically required more careful reading than you'd want. ai-toolkit doesn't make you do that.
None of this is surprising if you watch how FLUX landed. Black Forest Labs dropped something genuinely different and the tooling around it moved fast — faster than the old SD ecosystem had any reason to expect. The scripts that worked for SD 1.5 were not going to be the scripts that won for FLUX, and they weren't.
What's interesting is how quickly the standard calcified. It's been months. The question is already settled.
Counterpoints
Push back, extend the argument, or sharpen it. New counterpoints go through review before they show up here.
No approved counterpoints yet.