Diffusers Welcomes Flux-2 (huggingface.co)

Black Forest Labs’ new FLUX.2 is a ground-up second-generation image model now available in Hugging Face Diffusers — not a drop-in replacement for FLUX.1 but a distinct architecture optimized for multimodal diffusion. It retains the MM‑DiT + parallel DiT design (double‑stream blocks that join image/text only for attention, then single‑stream parallel blocks) but changes key internals: a single text encoder (Mistral 3 Small / Mistral Small 3.1) with 512 token max, shared timestep/guidance AdaLayerNorm‑Zero modulation across blocks, removal of bias parameters everywhere, fused projections (attention QKV fused with FF input) and a SwiGLU MLP. The block mix shifts heavily toward single‑stream blocks (FLUX.2 dev-32B: 8 double‑stream / 48 single‑stream), moving most parameters into the single‑stream path and introducing a new VAE plus resolution‑dependent timestep scheduling. FLUX.2 supports both text- and image-guided generation and can condition on up to 10 reference images. Practical significance: these design choices aim to improve throughput, scaling behavior and conditioning flexibility, but break compatibility with Flux.1 LoRAs and workflows. FLUX.2 is large (DiT + text encoder can exceed 80GB VRAM), yet Hugging Face provides multiple inference strategies — CPU offload (~62GB on an H100), FlashAttention3 for Hopper GPUs, 4‑bit bitsandbytes/NF4 quantization to run on ~24GB or ~20GB GPUs, remote text‑encoder endpoints and group_offloading down to ~8GB GPU (with higher host RAM). That mix of architectural changes plus accessible offloading/quantization paths makes FLUX.2 a technically interesting and more widely usable option for high‑fidelity image generation and fine‑tuning workflows.