Black Forest Labs (BFL) has released the FLUX 3 Image Klein model, designed for local image generation and editing. The core model features 5.6 billion parameters and supports image generation up to 4K resolution.
The model utilizes the Self-Flow system, developed in collaboration with MIT, which incorporates natural video data during training to improve spatial and temporal consistency in generated images.
The full pipeline requires several components: the Qwen3-VL-4B-Instruct text encoder (4.4 billion parameters), the FLUX 3 Image Klein core (5.6 billion parameters), and the flux2-vae.safetensors VAE (300 million parameters).
For optimal results, BFL recommends using the FLUX 3 Image Klein Prompt Upsampler, based on the Qwen3.5-9B model (9 billion parameters). This brings the total model size to approximately 20 billion parameters.
Testing on a system with an 8 GB GeForce 1070 GPU resulted in a 1.5-megapixel image generation time of 1 hour and 33 minutes when using the full pipeline within ComfyUI version 0.42.4.
