Credit: BlackForestLabs

Black Forest Labs has launched FLUX.2, its newest family of frontier image generation models, delivering major improvements in photorealism, consistency and text rendering. The models come with a new multi-reference capability, sharper typography, and support for detailed, pose-controlled image generation.

In collaboration with NVIDIA, FLUX.2 is available at launch in FP8 quantized variants that cut VRAM requirements by 40 percent while boosting performance by an equivalent 40 percent on RTX GPUs. The models run directly in ComfyUI, with no additional software required.

Next-generation visual quality

Credit: nvidia

FLUX.2 produces photorealistic images up to 4 megapixels, with real-world lighting and physics to eliminate the “AI look.” Artists can apply explicit pose control and generate clean, legible text across UI screens, infographics and multilingual layouts. The new multi-reference feature lets creators use up to six reference images to maintain consistent style and subjects without fine-tuning.

Built for RTX performance

The full FLUX.2 model is large, at 32 billion parameters, and requires 90GB VRAM to load fully or 64GB in low-VRAM mode — beyond the limits of typical consumer GPUs. To make the model widely accessible, NVIDIA and Black Forest Labs introduced FP8 quantization, reducing memory demand without a significant hit to quality.

NVIDIA and ComfyUI also enhanced weight streaming, enabling parts of the model to offload to system RAM, and added optimizations for FP8 checkpoints to further improve RTX performance.

Get started

Users can access FLUX.2 today by updating ComfyUI and exploring the built-in templates, or by downloading model weights from Black Forest Labs’ Hugging Face page.