Black Forest Labs just dropped the FLUX.2 family of image generation models, making it one of the most advanced visual AI releases yet, as well as a slate of new tools, including multi-reference generation that lets creators produce dozens of consistent image variations using up to six reference photos.
While life-like visuals are the bread and butter for image generation models these days, text on the other hand, is almost always the one element getting butchered. Therefore, FLUX.2 promises cleaner text rendering, better font handling at scale, and more control over subjects thanks to direct pose specification. All of this sits on top of a photorealistic engine capable of generating images at up to 4MP with realistic lighting and physics that help remove that typical AI sheen.
As for NVIDIA’s involvement in this, they worked closely with Black Forest Labs and ComfyUI to get FLUX.2 running with FP8 quantization and a range of GPU-level optimizations right out of the gate, cutting VRAM requirements by about 40% while boosting performance by roughly the same margin. Considering the full model weighs in at 32B parameters and typically demands 90GB of VRAM or 64GB even in low VRAM mode, these optimizations are crucial for bringing FLUX.2 anywhere near consumer RTX cards.
To make that practical, NVIDIA and ComfyUI also upgraded ComfyUI’s weight streaming system, which offloads parts of the model into system RAM. Well, it’s slower than native GPU memory for sure, but it dramatically extends the effective memory pool, making FLUX.2 usable on far more machines. Additionally, tuning for FP8 checkpoints has been optimized, too.
FLUX.2 is now available through ComfyUI’s downloadable templates, as well as Black Forest Labs’ Hugging Face page.
