Skip to content

FLUX.2 LoRA Inference, Explained Simply

What FLUX.2 LoRA Inference is, what it does well, and what it costs — released 2026-04-23 by Black Forest Labs.

TLDR
  • Black Forest Labs just opened a public beta for FLUX.2 LoRA Inference.
  • You can upload your own custom art styles (called LoRAs) and use them to generate images.
  • You don't need to rent or manage your own heavy-duty computer hardware to do it.
  • During the beta, using your custom styles costs the same as using the standard models.

Want a heads-up when FLUX updates?

We email you the moment a new FLUX version drops — plus plain-English release notes like these. No spam.

Choose update emails FLUX · Vibe Mastermind Updates
Choose which update emails you want

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.

If you’ve ever wanted an AI to draw pictures in your own specific art style, this new tool from Black Forest Labs is for you. It lets you upload a "LoRA"—a small file that teaches the AI a specific look or subject—and generate images through their system, without needing to be a computer hardware expert.

What it does well

This tool is all about customization. A "LoRA" (Low-Rank Adaptation) is basically a small add-on file that teaches a base AI model a new trick, like drawing a specific character or using a particular painting style. This update lets you bring a LoRA you trained yourself, upload it to their website dashboard, and start making images with it. It handles the heavy lifting on their end, so you don't have to worry about renting or maintaining expensive GPUs (the powerful computer chips needed for AI).

How it compares

This is a completely new, self-serve system that replaces Black Forest Labs' older FLUX.1-era Finetuning API, which was shut down in October 2025. Unlike the old way, this new path lets you train your custom files using whatever outside tools you already like (such as AI-Toolkit or Diffusers) and just upload the finished product.

What it costs and what it can handle

During the public beta, using your custom LoRA costs the exact same amount as using the regular, un-customized base models. It gives you six different "endpoints" (the specific addresses your software talks to) to choose from. These cover different sizes and versions of the FLUX.2 [klein] model, including 4B and 9B options (the numbers refer to the model's size in billions of parameters), as well as different data formats like FP8 and BF16. You can also adjust the "finetune_strength" to control exactly how strongly your custom style shows up in the final image.

Worth knowing

You need to use their API to make this work, meaning it’s built for people who are somewhat comfortable hooking up code, not just a simple click-and-type website. Also, while you can use outside tools to train your LoRA, you have to upload it through their specific Dashboard before you can use it.

This is our plain-language summary. Read the complete, official notes on the FLUX official changelog ↗.

Vibe Mastermind in 00d 00h 00m 00s
Call LIVE — Join Now