Black Forest Labs — the frontier AI research lab developing visual generative AI models — today released the FLUX.2 family of state-of-the-art image generation models.
FLUX.2 is packed with new tools and capabilities, including a multi-reference feature that can generate dozens of similar image variations, in photorealistic detail and with cleaner fonts — even at scale.
NVIDIA has worked with Black Forest Labs and ComfyUI to make the models available with FP8 quantizations and RTX GPU performance optimizations at launch, decreasing the VRAM required to run them by 40% and improving performance by 40%.
Requiring no special software package to run, the models are available directly in ComfyUI.
State-of-the-Art Visual Intelligence
Images generated by FLUX.2 are photorealistic, even at scale, featuring up to 4 megapixel resolution with real-world lighting and physics to eliminate that “AI look” that undermines visual fidelity.
The models add direct pose control to explicitly specify the pose of a subject or character in an image, as well as deliver clean, readable text across infographics, user interface screens and even multilingual content. Plus, the new multi-reference feature enables artists to select up to six reference images where the style or subject stays consistent — eliminating the need for extensive model fine-tuning.
For a complete overview of new FLUX.2 features, read Black Forest Labs’ blog.
Optimized for RTX
The new FLUX.2 models are impressive, but also quite demanding. They run a staggering 32-billion-parameter model requiring 90GB VRAM to load completely. Even using lowVRAM mode — a popular setting that allows artists to only load the active model at a time — the VRAM requirement is still 64GB, which puts the model virtually out of reach for any consumer card to use effectively.
To broaden FLUX.2 model accessibility, NVIDIA and Black Forest Labs collaborated to quantize the model to FP8 — reducing the VRAM requirements by 40% at comparable quality.
And to make this model accessible on GeForce RTX GPUs, NVIDIA has partnered with ComfyUI — a popular application to run visual generative AI models on PC — to improve the app’s RAM offload feature, known as weight streaming.
Using the upgraded feature, users can offload parts of the model to system memory, extending the available memory on their GPUs — albeit with some performance loss, as system memory is slower than GPU memory.
NVIDIA has also been collaborating with ComfyUI to optimize model performance on NVIDIA and GeForce RTX GPUs, including optimizations for FP8 checkpoints.
Get started with FLUX.2 today. Update ComfyUI and check out the FLUX.2 templates, or visit Black Forest Labs’ Hugging Face page to download the model weights.
Plug in to NVIDIA AI PC on Facebook, Instagram, TikTok and X — and stay informed by subscribing to the RTX AI PC newsletter.