FLUX

FLUX is a family of text-to-image and image-editing generative models developed by Black Forest Labs, a German research company founded in 2024 by former Stability AI researchers. First released on 1 August 2024, FLUX models became known for prompt adherence, photorealistic detail, legible text rendered inside images, and a distribution strategy that pairs commercial APIs with openly downloadable model weights.^[1][2]

History

Black Forest Labs was founded in 2024 by Robin Rombach, Andreas Blattmann, Patrick Esser, Dominik Lorenz and colleagues — researchers associated with the latent diffusion line of work behind Stable Diffusion. The company launched the first FLUX family on 1 August 2024 in three tiers: FLUX.1 [pro], available through an API; FLUX.1 [dev], released as open weights under a non-commercial licence; and FLUX.1 [schnell], a 12-billion-parameter model under the permissive Apache 2.0 licence that generates images in only one to four sampling steps. An improved FLUX.1.1 [pro] followed in October 2024, and FLUX.1 Kontext, a model for reference-guided image editing, arrived in 2025 alongside FLUX.1 Krea, an aesthetics-focused collaboration with the creative tool company Krea. In November 2025 Black Forest Labs released FLUX.2, headlined by FLUX.2 [dev], a 32-billion-parameter open-weight checkpoint that combines text-to-image synthesis with multi-reference editing; the compact [klein] family followed in January 2026, including a 4-billion-parameter model under Apache 2.0 that runs on consumer GPUs. In December 2025 the company announced a US$300 million Series B round at a US$3.25 billion valuation, with investors including Salesforce Ventures, Andreessen Horowitz, NVIDIA and Temasek.^{[1][3][4][5][6]}

Key Concepts and Technology

FLUX models are rectified flow transformers: diffusion-based networks that learn a velocity field transporting random noise into image latents, conditioned on text by large language-model-style encoders. The [schnell] variant is distilled with latent adversarial diffusion distillation so that usable images emerge in a handful of steps, while the flagship tiers trade speed for fidelity and controllability. FLUX.2 added multi-reference editing — combining several input images, subjects and styles in a single generation — at up to four-megapixel resolution, along with stronger world knowledge and text rendering. Because the [dev] and [klein] tiers ship as open weights, the family has a broad community ecosystem: workflows in ComfyUI, quantised builds for consumer hardware, an fp8 implementation optimised for NVIDIA RTX GPUs, LoRA fine-tunes for custom styles, and hosted endpoints from services such as fal, Replicate, Together AI and Cloudflare. Black Forest Labs has not published the full composition of its training data, and the non-commercial [dev] licence requires a commercial agreement for business deployment.^[2][3][7][8]

Applications and Impact

FLUX is used across professional creative workflows: advertising and brand campaigns, product visualisation, concept art, storyboarding, and design mock-ups, with the open weights making it a default choice for local, self-hosted generation and for fine-tuned vertical models. The family competes with closed systems such as Midjourney, Adobe Firefly and Google's image models, and with other open-weight alternatives, in a market where quality, licensing and control over data are the main differentiators. Its release cycle has also fed industry debate over training data, copyright and the labelling of AI-generated media.^[3][8]

>See Also

• Stable Diffusion
• ComfyUI
• Midjourney
• Text-to-Image Model
• Open-Weight Models