Black Forest Labs released the FLUX.2 family of state-of-the-art visual generative Artificial Intelligence models. The models introduce a multi-reference feature that can generate dozens of similar image variations while maintaining consistent style, plus direct pose control to explicitly specify a subject or character pose. FLUX.2 emphasizes photorealism and fidelity, producing images with up to 4 megapixel resolution, real-world lighting and physics, and cleaner, more readable text suitable for infographics, user interface screens and multilingual content.
FLUX.2 models are large and demanding. The family includes a 32-billion-parameter model that requires 90GB VRAM to load completely and still needs 64GB in lowVRAM mode, placing the model beyond the reach of most consumer cards. To broaden accessibility, NVIDIA and Black Forest Labs collaborated to quantize the model to FP8, a change that decreases the VRAM required to run the models by 40% at comparable quality and also produced improvements in performance by 40% in NVIDIA tests. NVIDIA has additionally worked with ComfyUI to deliver RTX GPU performance optimizations and FP8 checkpoint support. ComfyUI integration requires no special package to run and offers templates that make it straightforward to use FLUX.2 on desktop systems.
To enable use on GeForce RTX GPUs with more modest memory, ComfyUI’s RAM offload feature, known as weight streaming, was improved so users can offload parts of the model to system memory, extending usable capacity at the cost of some performance due to slower system memory. The models support artist workflows such as selecting up to six reference images to keep subject or style consistent without fine-tuning. Black Forest Labs provides a full feature overview on its blog, and model weights are available on the lab’s Hugging Face page. Users are advised to update ComfyUI to access FLUX.2 templates and the optimized FP8 checkpoints for NVIDIA and GeForce RTX platforms.
