Flux: The Next Generation
Overview
Flux is an image generation model by Black Forest Labs, the founding team behind Stable Diffusion. Released mid-2024, it represents a major quality leap over SDXL.
Flux Variants
- Flux.1 Dev: non-commercial, full version
- Flux.1 Schnell: Apache 2.0, fast (4 steps)
- Flux.1 Pro: API only, commercial
Why Flux is Superior
- Readable text in images
- Better prompt following
- Photorealistic quality
- Better human anatomy
- Superior composition understanding
ComfyUI Configuration
Flux uses separate model loading: Load Diffusion Model + DualCLIPLoader + Load VAE.
Recommended Parameters
- sampler: euler, scheduler: simple
- Flux Dev: 20-30 steps, Flux Schnell: 4 steps
- cfg: 1.0 (guidance via FluxGuidance node, typically 3.0-4.0)
VRAM Optimization
- Use quantized models (FP8, NF4/GGUF)
- T5 encoder in FP8
--lowvramflag- CPU offloading
LoRAs, ControlNet, and img2img
All work similarly to SD but with Flux-specific models and configurations. LoRAs are larger (~100-500 MB). ControlNet support is newer but growing.