Flux.1
Flux.1 by Black Forest Labs: a large image model (dev, schnell and tool models) that reads sentences and follows prompts closely.
What it is
Released in 2024, Flux.1 is a 12-billion-parameter transformer. dev is the quality version (non-commercial licence for the weights); schnell is distilled for 4 steps (Apache 2.0). Black Forest Labs also published Fill (inpainting and outpainting), Canny and Depth (control), Redux (image variations) and Kontext (editing by instruction). A large ecosystem of fine-tunes and LoRAs exists for dev.
Files and nodes
- Load Diffusion Model:
flux1-devorflux1-schnell(fp8 versions about 12 GB) inmodels/diffusion_models. Some all-in-one fp8 checkpoints load with Load Checkpoint instead. - DualCLIPLoader:
clip_landt5xxl(fp16 or fp8),typeflux. - Load VAE:
ae.safetensors. - EmptySD3LatentImage; FluxGuidance on the positive conditioning; often ModelSamplingFlux; many templates use SamplerCustomAdvanced with BasicGuider.
Settings that work
About 1 megapixel, flexible up to about 2; dev: 20 to 30 steps, cfg 1, FluxGuidance about 3.5 (lower for more realism), euler with simple or beta. schnell: 4 steps, cfg 1. Prompt: natural sentences; it can write short text in quotes. The negative prompt does nothing at cfg 1.
What fits
Only Flux.1 LoRAs and ControlNets. Flux.2 files are not compatible. FLUX.1 Krea [dev] is a Flux.1 dev model trained again by Krea: it loads exactly like Flux.1, and Flux.1 dev LoRAs usually load, though results can differ. Chroma is related but behaves differently with Flux.1 LoRAs. Krea 2 is another architecture: Flux files do not work with it.
How the app helps
Shown as Flux.1 on model cards. The built-in knowledge describes FluxGuidance and ModelSamplingFlux. The lessons Flux.1: a picture from separate files, A quantised model: GGUF, A reference picture for style (Flux Redux) and The same character in a new picture: FLUX.1 Kontext work with Flux.1.