AI BriefingBlack Forest Labs Unveils FLUX.2
·2025.11.25 09:00
Key point
Black Forest Labs' new image generation model FLUX.2 has been released and is available in the Diffusers library.
1 / 2
Details
Black Forest Labs has unveiled FLUX.2, a new image generation and editing model. Rather than a simple update to FLUX.1, it is a completely new architecture pre-trained from scratch.
Key Architectural Changes
- Text Encoder: Moving away from the existing dual-encoder structure, it adopts a single Mistral Small 3.1 encoder, simplifying prompt embedding computation.
- Improved DiT Structure: Bias parameters have been removed from all layers, and it is designed so that time and guidance information are shared across all blocks.
- Parallel Transformer Block: A fully parallel transformer block was introduced, fusing the Attention QKV projection and the FF input projection.
- Expanded Single-Stream Ratio: The proportion of single-stream blocks among total parameters has been significantly increased compared to FLUX.1.
Inference and Usage
- VRAM Requirements: As the model scale has grown, inference without offloading requires 80GB or more of VRAM.
- Library Support: Image generation, editing, and LoRA fine-tuning are possible via Hugging Face's Diffusers library.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.