NVIDIA and Black Forest Labs Just Dropped a Monster Image Model
FLUX.2 is here, it is huge, and it wants every RTX GPU it can get its hands on.
Black Forest Labs has unveiled FLUX.2, a new family of image generation models that looks purpose-built for people who think their GPUs are not working hard enough.
The models promise sharper detail, faster generation and more creative control. They also arrive with a polite warning: they are hungry.
The headline feature is multi-reference generation. Feed the model a handful of source images and it will spin out photorealistic variations that look like they were shot by the same camera on the same day.
It is the kind of tool artists have been hacking together with plugins for years. Now it is built straight into the model’s core.
NVIDIA has been deeply involved in the launch. The two companies worked together to optimise FLUX.2 for RTX hardware using FP8 quantisation. The result is a striking set of gains. VRAM requirements drop by about 40 percent.
Performance jumps by about the same margin. Black Forest Labs says the models are impressive but demanding, which is a polite way of saying they will still melt lesser GPUs.
The full version of FLUX.2 weighs in at 32 billion parameters and wants 90 GB of VRAM to run in its natural state.
Low-VRAM mode brings that down to 64 GB. It is still heavy enough that most creators will rely on NVIDIA’s other big contribution: ComfyUI optimisation. Users can now offload chunks of the model to system memory, which helps keep RTX cards within shouting distance of the required capacity.
Anyone who wants to experiment can update ComfyUI and load the FLUX.2 templates. Model weights are available on the Black Forest Labs page on Hugging Face.
FLUX.2 is a reminder of where cutting-edge visual models are heading. Bigger, more capable and more computationally intense.
Black Forest Labs says it is building the frontier of generative imaging. With NVIDIA’s help, it may have just pushed that frontier a few notches forward.
