FLUX.2
Next-generation image generation model by Black Forest Labs.
About
FLUX.2 is Black Forest Labs' second-generation image generation and editing family, succeeding FLUX.1 with a reworked architecture: a latent flow matching design that couples a 24B Mistral vision-language model with a rectified flow transformer and a newly redesigned VAE. The models generate and edit images up to 4 megapixels, follow complex structured prompts more reliably, render legible typography for infographics and UI mockups, and support multi-reference editing that keeps characters and products consistent across up to 10 reference images. The lineup spans hosted variants, pro and flex, offered through the BFL API and partner platforms, alongside open-weight releases: FLUX.2 dev, a 32B checkpoint distributed through Hugging Face for local use, and klein, a smaller Apache 2.0 licensed variant. Typical users include professional designers, creative tooling companies, and developers who run the open weights locally or through API platforms such as FAL, Replicate, and Cloudflare.
Should you use FLUX.2?
Pick it when
Pick FLUX.2 when you need editing that keeps a character or product consistent across up to 10 reference images, legible typography for infographics or UI mockups, or 4MP output, and you can host a 32B model or use the API.
Look elsewhere when
Skip the 32B dev checkpoint on a single consumer GPU: FLUX.1, Z-Image, or Mage-Flow are far lighter. For commercial use of open weights, klein is the Apache 2.0 variant, so check the dev terms or pick Qwen-Image.
Alternatives to FLUX.2
- FLUX.1
Much lighter than the 32B dev checkpoint and runs on 12 GB GPUs with mature frontend support and companion editions, but lacks multi-reference editing and 4MP output.
- Qwen-Image
Apache-2.0 20B model with paragraph-level English and Chinese text and an edit line with multi-image input, smaller than the 32B dev checkpoint.
- Mage-Flow
MIT-licensed 4B model that fits one workstation GPU and, per its authors, scores 0.90 on GenEval versus 0.87 for FLUX.2-dev, but with a much younger ecosystem.
- Z-Image
Apache-2.0 6B model with bilingual text rendering and a Turbo variant on 16 GB, trading multi-reference editing for far lower hardware cost.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- Image Generation
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- Minimum VRAM
- 12 GB
- Added
- Apr 3, 2026
Tags
Mentioned in
Open Source Alternatives to AI SaaS in 2026: 8 Honest Swaps
Eight hosted AI products, the open source and open-weight tools that replace them, the license and hardware...
BuilderAI Editorial Team
The Best Open-Source Image Generation Stack in 2026
Model picks per use case, the UI that actually won, a working upscaling and editing chain, and an honest read...
Max P
From ComfyUI to Production: Serving Diffusion Models in 2026
A working guide to moving diffusion pipelines from ComfyUI prototype to production: environment locking,...
Max P