FLUX.2

FLUX.2 is the current generation of open and API text-to-image models from Black Forest Labs, the team of former Stable Diffusion researchers behind the original FLUX.1. Where the first FLUX.1 paired its transformer with a T5 text encoder, FLUX.2 moves to the Mistral Small 3.2 text encoder, giving it noticeably stronger language understanding — it parses long, natural-language descriptions and handles instructions in multiple languages. The result is sharper prompt following, cleaner in-image text, and more believable photorealism than the model it replaces.

The lineup is a set of tiers rather than a single model. FLUX.2 [dev] is the open-weight, quality-first checkpoint (weights are published on Hugging Face and hosted here on Civitai); the FLUX.2 [klein] family — released in 9B and 4B sizes, plus matching "base" variants — is Black Forest Labs’ fastest line yet, folding generation and editing into one compact architecture that can finish an image in roughly a second. Above the open weights sit the API-only tiers — Pro and Flex — tuned for higher fidelity and finer control. Every one of them is generatable on Civitai, so you can move between the open [klein] and [dev] checkpoints and the hosted API tiers without downloading weights or standing up a local rig.

Choose FLUX.2 when prompt fidelity, legible typography, precise color, or photographic realism matter most — product shots, posters, text-in-image work, and complex multi-subject scenes are where it shines. If you want the smallest, fastest footprint for local iteration, reach for a [klein] variant; if you want maximum quality and are generating on Civitai anyway, [dev] or the API tiers give you more headroom. For anime and character art the SDXL-based Pony and Illustrious ecosystems still hold a deeper LoRA library, but FLUX.2’s own LoRA catalog is growing quickly on top of the open [klein] and [dev] bases.