Overview
- The family spans managed tiers (pro, flex) available through BFL Playground and APIs, plus a 32B open-weight dev checkpoint downloadable on Hugging Face.
- Core upgrades include photoreal output up to 4MP, cleaner typography, direct pose control, stronger prompt following, and multi-reference inputs for consistent characters, products or styles.
- NVIDIA reports the full model needs about 90GB of VRAM, or roughly 64GB in low-VRAM mode, underscoring the reliance on quantization, offload strategies or hosted endpoints for many users.
- NVIDIA’s FP8 checkpoints reduce memory demands and improve throughput by around 40%, enabling more practical runs on GeForce RTX GPUs at comparable quality.
- ComfyUI adds native templates plus weight streaming to offload parameters to system memory, offering broader access with some performance trade-offs.