FLUX.2 Released: Black Forest Labs' 32B-Parameter Open Image Model

Black Forest Labs released the FLUX.2 image generation system on November 25, 2025 — its most powerful open image generation and editing model to date, squaring off against leading closed products like Google's Nano Banana.

FLUX.2-dev is a 32-billion-parameter rectified flow transformer capable of generating, editing and combining images from text. Key upgrades include multi-reference capability — conditioning on multiple input images at once to keep characters and styles consistent — built for real creative workflows rather than demos. The weights are available on Hugging Face.

The Baton After Stable Diffusion#

BFL was founded by the original Stable Diffusion team, and after SD 3.5 the open-image community's center of gravity had already shifted to the FLUX line: default base models, LoRA fine-tunes and workflow templates increasingly build on FLUX. Pushing the open flagship to 32B parameters amounts to a declaration: the technical ceiling of open image generation is ours to define (see our image tools review).

Multi-Reference: Matching the Closed Camp's Key Capability#

Nano Banana broke out on character consistency (see our coverage); FLUX.2's multi-reference conditioning is the open camp's answer to exactly that capability — anchoring person, product and style at once, making brand assets and series work genuinely usable. The open-vs-closed image fight has moved from single-shot beauty into the workflow-consistency half.

The Business Behind Open Weights#

Open weights are not charity: the dev build serves community and research, while commercial licenses and flagship API tiers earn the revenue — the same open-for-reach, closed-for-monetization dual-track as gpt-oss and Llama (see our coverage). For an independent lab like BFL, the open-source ecosystem position is the only moat against the giants' distribution advantage.

Our Take#

ComfyUI and other mainstream tools integrated FLUX.2 quickly, upgrading the whole open image stack. Its sample lesson: in the same year a closed viral hit (Nano Banana) captured mass mindshare, the open camp held the professional-workflow base — demand for controllable, fine-tunable, self-hostable image models does not disappear; it grows with compliance requirements (see our local deployment review).