Black-forest-labs/FLUX.2-dev (huggingface.co)

Black Forest Labs has open‑released FLUX.2 [dev], a 32‑billion‑parameter "rectified flow" transformer for text-to-image generation, single- and multi-reference editing, and image fusion. The model requires no additional fine‑tuning to adopt character, object, or style references and was trained with guidance distillation to improve inference efficiency. The team publishes an official diffusers pipeline, sampling code, and a 4‑bit quantized example for consumer GPUs (e.g., RTX 4090) that can use a remote text encoder; a commercial FLUX.2 Pro API is also available. This positions FLUX.2 as one of the most capable open-weight T2I/I2I systems for researchers and creatives who want editable, multi-reference synthesis without retraining. Black Forest Labs emphasizes safety and responsible release: pre‑training data filtering (with the Internet Watch Foundation), targeted post‑training fine‑tuning to suppress CSAM/NCII generation, multiple third‑party adversarial evaluations (≈2,800 prompts) that reportedly show higher resilience than other open models, and runtime input/output filters (including third‑party vendors). The repo implements pixel-layer watermarking and links to C2PA provenance; API outputs carry cryptographically signed C2PA metadata. The open weights are available under a FLUX [dev] Non‑Commercial License and require agreeing to access terms and contact disclosure, while the commercial API provides additional guarded deployment options.