Because full BF16 FLUX.1 weights consume ~24 GB VRAM, community FP8 and GGUF quantized Checkpoints quickly unlocked 8–12 GB consumer GPUs, while Fal.ai and Replicate offered sub-second hosted FLUX endpoints.
FLUX
Also known as: FLUX.1 · FLUX.1-dev · FLUX.1-schnell · FLUX.1 Kontext
Plain English
A family of open-weight and API image generation models from Black Forest Labs that replaced Stable Diffusion XL as the community's go-to base model for sharp photorealism, accurate text in images, and custom LoRA training.
Technical Overview
FLUX.1 is a 12B-parameter rectified-flow multimodal DiT paired with a 16-channel VAE and dual CLIP + T5-XXL text encoders. Its dev checkpoint uses guidance Distillation & Synthetic Data (replacing dual-pass CFG), while schnell uses step distillation for 1–4 step inference—widely quantized to FP8 and GGUF for ComfyUI.
Chronological Entries (2)
Black Forest Labs released the 12B FLUX.1 family (pro, dev, schnell) on Hugging Face. Within weeks it became the primary target for LoRA & QLoRA fine-tuning on Civitai and modular workflows in ComfyUI.
Direct Relations (4)
- Created byBlack Forest Labs
- ImplementsDiffusion Transformer (DiT)
- UsesLatent Space & VAE
- Part ofComfyUI