FLUX.2 Dev

The model weights are released under the FLUX [dev] Non-Commercial License v2.0, which restricts direct commercial deployment of the weights without a separate license. However, output images can be used commercially (you just cannot resell or redistribute the model itself without a license). For commercial deployment, BFL offers a paid Self-Hosted Commercial License. Using FLUX.2 Dev through licensed API providers (including Masonry) includes commercial-use output rights in their service terms.

At full BF16/FP16 precision the 32-billion-parameter model needs roughly 80GB of VRAM, so a data-center GPU (such as an H100 or A100) or a multi-GPU setup is required. The NVIDIA-optimized FP8 build cuts that to around 24GB, which fits an RTX 4090 or 3090, the practical minimum for comfortable single-card inference. 4-bit GGUF quantization reduces it further to roughly 12–16GB on mid-range consumer cards, with modest quality trade-offs. You also need at least 32GB of system RAM and around 60GB of storage for the weights, VAE, and text encoders.

FLUX.2 Dev ships with open weights you can download, self-host, and fine-tune. FLUX.2 Flex is API-only and exposes inference parameters (steps up to 50, guidance scale) that let you tune the speed-quality trade-off at generation time. Flex also has a stronger typography specialization for text-heavy creative. The choice is architectural: Dev if you need to own the model and customize it; Flex if you want parameter-level control through an API without managing infrastructure.

Yes. FLUX.2 Dev supports LoRA (Low-Rank Adaptation) fine-tuning. A typical brand LoRA trains on 20–50 high-quality images representing your visual style, takes 2–8 hours on a consumer GPU, and produces a 50–200MB adapter file. The adapter loads alongside the base checkpoint at inference without modifying the original weights, so you can switch between base and fine-tuned outputs without re-deploying. This is one of the primary reasons development teams choose Dev over API-only models.

FLUX.2 Pro is the closed-source production API variant optimized for maximum quality and reliability. FLUX.2 Dev is a close second in output quality, close enough that for most creative briefs the difference is minor. The meaningful differences are in the deployment model (open weights vs. API-only) and fine-tuning capability (Dev supports it; Pro does not). Teams that need the absolute ceiling of FLUX.2 image quality and do not need custom weights typically choose Pro.

Yes. The unified architecture means a single FLUX.2 Dev checkpoint handles both text-to-image generation and image editing without switching models or adding separate modules. You can use it for targeted edits (changing backgrounds, modifying elements, refining details), which simplifies both self-hosted deployment and the prompting workflow since you stay in one context rather than switching tools.

Rectified flow is a training and sampling approach that learns to map noise to images along straighter trajectories than traditional diffusion, which tends to produce cleaner results with fewer inference steps. Combined with the Transformer backbone (as opposed to the U-Net used in earlier diffusion models), it gives FLUX.2 Dev strong compositional ability and efficient scaling. In practice this means fewer steps needed for quality results and better handling of complex multi-element prompts.

Up to 10 reference images natively, without requiring external ControlNet modules or additional pipeline components. This makes it practical for brand consistency workflows where you need to simultaneously reference a product, a visual style, a character, and a composition, all in a single inference call rather than a multi-step pipeline.

Yes, and this is a primary use case. Self-hosting means your prompts, reference images, and generated outputs never leave your infrastructure. With the BFL Self-Hosted Commercial License (available for purchase at bfl.ai/pricing/licensing), you can deploy FLUX.2 Dev in your own cloud or on-premises environment and integrate it into internal tools or client-facing products. You cannot, however, redistribute the model weights or expose them through a resale API without BFL's explicit authorization.

Yes. Black Forest Labs provides reference inference code, and optimized FP8 implementations exist for ComfyUI and other community tooling. The community has also produced numerous GGUF quantizations for reduced-VRAM deployments. If you prefer a managed path without local setup, Masonry provides access to FLUX.2 Dev through the platform without requiring GPU infrastructure on your side.

FLUX.2 Dev handles short-form text in images well (product labels, packaging copy, brand names), which is meaningful for marketing creative. For complex typography, infographics, and dense multi-line text layouts, FLUX.2 Flex has a stronger specialization. GPT Image 2 remains the benchmark for dense text and small lettering. The right choice depends on your brief: Dev for general photorealistic creative with moderate text; Flex or GPT Image 2 when text rendering is the primary requirement.

FLUX.2 Dev is an AI image generation model from Black Forest Labs, available inside Masonry, the AI creative agent teams use to produce marketing, product, and brand images.

Open a Masonry canvas, pick FLUX.2 Dev from the model selector, and describe the image you need: a product shot, an ad creative, a social post. Masonry generates it, then you refine, edit, and combine FLUX.2 Dev with other models in one workspace.

Yes, you can start generating images with FLUX.2 Dev on Masonry's free tier, then scale up with higher limits and priority processing as your team grows.

FLUX.2 Dev is built by Black Forest Labs. Inside Masonry it runs alongside 50+ image and video models, so your team can pick the right one for each brief without switching tools.

Yes, the prompt gallery on this page shows real images teams have generated with FLUX.2 Dev in Masonry, each paired with the exact prompt you can copy and adapt for your own brand.