Diffusers welcomes FLUX-2

FLUX.2 is a new open image generation model from Black Forest Labs, featuring a novel architecture for both text-guided and image-guided generation. It utilizes a single Mistral Small 3.1 text encoder and a multimodal Diffusion Transformer (MM-DiT) architecture with significant changes from its predecessor. Key architectural updates include shared time and guidance information across transformer blocks, the removal of bias parameters, and the use of fully parallel transformer blocks. The model requires significant VRAM for standard inference, but methods for running on resource-constrained systems using 4-bit quantization and CPU offloading are provided, along with instructions for LoRA fine-tuning.