Black Forest Labs launches FLUX.1 with $31M backing

The original minds behind Stable Diffusion are officially back, and they are not wasting any time.

After departing Stability AI, the core research team has launched Black Forest Labs, backed by a fresh $31 million Seed round led by Andreessen Horowitz. Alongside the funding, the European startup dropped FLUX.1, a brand-new family of text-to-image models that directly challenge Midjourney and OpenAI.

Three models for three different crowds

The team behind VQGAN, Latent Diffusion, and SDXL rolled out three distinct flavors of FLUX.1 right out of the gate.

At the top sits FLUX.1 [pro]. This is the flagship, closed-access model designed for heavy commercial workloads, accessible through the company's API as well as platforms like fal.ai and Replicate.

In the middle is FLUX.1 [dev], an open-weight model directly distilled from the pro tier. It preserves high prompt adherence and detail while running far more efficiently, though it is restricted to non-commercial research.

Then comes FLUX.1 [schnell], German for fast. This model is built for local development, runs in very few steps, and arrives under a completely open Apache 2.0 license. Hugging Face Diffusers and ComfyUI supported it on day one, meaning local enthusiasts can spin it up immediately.

A 12-billion parameter shift to flow matching

Under the hood, FLUX.1 is a massive departure from older generation pipelines. All public variants scale up to 12 billion parameters, leveraging a hybrid architecture of multimodal and parallel diffusion transformer blocks.

Instead of conventional diffusion training, the team built on flow matching, incorporating rotary positional embeddings and parallel attention to push hardware efficiency. The result is a model that native-handles resolutions from 0.1 to 2.0 megapixels across flexible aspect ratios.

Crucially, Black Forest Labs claims FLUX.1 outperforms Midjourney v6.0, DALL-E 3, and SD3-Ultra in raw visual fidelity, complex prompt following, and especially typography, rendering crisp in-image text without the typical AI gibberish. Next up, the team is already teasing a state-of-the-art text-to-video suite.

Why it matters

We think this is the most critical reset the open generative media community has seen all year. Following community disappointment over SD3's licensing and prompt handling, many feared open-source image generation was losing ground to proprietary giants like Midjourney.

For creators and developers, FLUX.1 offers an immediate, practical upgrade. Having an Apache 2.0 licensed model like schnell integrated into ComfyUI on day one means indie developers can build local pipelines without worrying about surprise license rug-pulls. If you need pure fidelity for client work, the pro API provides an accessible drop-in replacement for closed alternatives.

Tactically, Black Forest Labs played this brilliantly. By splitting the lineup between a monetization-ready enterprise tier and fully open weights for the community, they capture commercial revenue while instantly winning back the grassroots developer mindshare. The main catch will be hardware: a 12B parameter model is hefty, and running the higher-precision dev variant locally will stretch consumer GPUs. Still, the bar for visual fidelity and accurate text rendering has officially been raised.