Flux.2
Flux.2 by Black Forest Labs: the second generation of Flux, for pictures and for editing with several reference pictures.
What it is
Flux.2 (late 2025) is a new architecture, not an update of Flux.1 files. Its dev model is large and uses a big language model as its text encoder (a Mistral model); smaller Klein models followed for ordinary cards. Flux.2 can take reference pictures and edit or combine them by instruction.
Files and nodes
Load Diffusion Model with the Flux.2 model; Load CLIP with its own text encoder, type flux2; Load VAE with the Flux.2 VAE (flux2-vae); Empty Flux 2 Latent (EmptyFlux2LatentImage), the empty latent of the official template; reference pictures go into the conditioning through reference nodes. dev and Klein use different text encoders: dev the Mistral model mistral_3_small_flux2, Klein 4B qwen_3_4b and Klein 9B qwen_3_8b; all share flux2-vae. The Klein graph of the templates builds its sampler from parts: Flux2Scheduler, CFGGuider, RandomNoise, KSamplerSelect and SamplerCustomAdvanced. Start from ComfyUI's Flux.2 or Flux.2 Klein template for the exact files.
Settings that work
Take steps, guidance and sizes from the template: dev works around 1 to 4 megapixels with guidance; distilled Klein versions use few steps and cfg 1. Prompt: detailed sentences; it handles text in pictures and structured prompts.
What fits
Only Flux.2 LoRAs, text encoder and VAE. Flux.1 files (LoRAs, t5xxl, ae.safetensors) do not work with it.
How the app helps
Shown as Flux.2 on model cards. The built-in knowledge names both text encoders: the Mistral-based one for dev, qwen_3_4b or qwen_3_8b for Klein. The lesson Flux.2 Klein: a sampler built from parts builds the Klein graph and checks its files.