Black Forest Labs released FLUX.2, an update to their family of image generation models designed for production workflows, with GPU optimizations developed alongside NVIDIA and ComfyUI that cut VRAM requirements by 40% and improve performance by 40%.
Multi-reference consistency - Reference up to 10 images simultaneously to maintain character, product, or style consistency across generated variations. Black Forest Labs claims this delivers "the best character / product / style consistency available today."
Professional text rendering - Generates legible typography in complex infographics, UI mockups, and multilingual content. The model can edit images at resolutions up to 4 megapixels while preserving detail.
32-billion-parameter model on consumer hardware - FLUX.2 [dev] normally requires 90GB VRAM to load completely, but FP8 quantization developed with NVIDIA reduces that by 40%. ComfyUI's upgraded weight streaming feature offloads parts of the model to system RAM, extending available GPU memory on consumer RTX cards.
Open-weight and API options - FLUX.2 [dev] weights are available on Hugging Face with reference inference code on GitHub. Production APIs available through BFL's API, plus endpoints on FAL, Replicate, Runware, Together AI, Cloudflare, and DeepInfra.
Built on Mistral-3 VLM - The architecture combines the Mistral-3 24B parameter vision-language model with a rectified flow transformer, bringing real-world knowledge and spatial reasoning that earlier architectures couldn't render.
This marks the first time a 32-billion-parameter image generation model launches with optimization specifically targeting consumer-grade workstations. The combination of FP8 quantization and improved memory offloading makes near-enterprise performance accessible on high-end desktop GPUs—relevant for concept artists, visual development teams, and anyone building asset pipelines that need stylistic consistency without manual fine-tuning.
