Black Forest Labs released FLUX Pro 2 yesterday — the second generation of the image diffusion model line that originated from the Stable Diffusion creators. On a 5,000-prompt blind human-preference evaluation run by Artificial Analysis, FLUX Pro 2 wins 53% of comparisons against Midjourney v7, becoming the first open-source-licensed image model to cross 50%.
The headline numbers tell the story: 1024x1024 generation in 1.8 seconds on an H100, native 4K output mode, full prompt adherence at long instruction lengths, and CC-BY-SA commercial use for the open-weights release.
What's actually new in FLUX Pro 2
Three architectural changes drove the quality jump:
- **Rectified Flow Transformer 2 backbone**: 18 billion parameters (up from 12B in FLUX 1.1)
- **Dual text encoder**: T5-XXL + CLIP-G fused with cross-attention, removes the prompt-bleeding issues of single-encoder models
- **Native multi-aspect training**: model trained on 1:1, 16:9, 9:16, 21:9, and 1:2 simultaneously — no quality drop on non-square output
- **Photorealism + style separation**: a single style-control vector lets you slide from "photorealistic" to "illustrated" continuously without retraining
What it solves that prior models couldn't:
- **Text rendering**: legible English/CJK/European at 8-12 word phrase level (FLUX 1.1 maxed at 4-6 words)
- **Anatomy at distance**: hands, fingers, eyes correct even at 1/8 frame size
- **Long prompt adherence**: 200+ word prompts execute correctly without entity confusion
Three pricing tiers
- **FLUX Pro 2** (closed, fastest): API at $0.04 per megapixel, exclusive to BFL platform
- **FLUX Dev 2** (open weights, non-commercial): free download, ~25 GB
- **FLUX Schnell 2** (open weights, commercial OK): free download, 4-step distilled version, ~12 GB
For builders: FLUX Schnell 2 is the killer release. It ships under Apache-2.0 license, runs on a single RTX 4090 with 24GB VRAM, and produces output indistinguishable from Pro for most use cases. The 4-step distillation makes inference 8x faster than FLUX 1.1 at comparable quality.
What this means for the field
For a year, the consensus has been "Midjourney v7 is the quality leader, FLUX is the open alternative". FLUX Pro 2 collapses that gap. Midjourney's moat narrows to brand and Discord workflow.
For Stability AI: bad week. They were already losing relevance after the Stable Diffusion 3 reception. FLUX Pro 2 takes the open-source crown decisively.
For builders running production image gen: switch tests are essentially mandatory. FLUX Schnell 2 at $0/inference (self-hosted) plus Pro 2 at $0.04/megapixel for premium fallback covers most workloads at a fraction of the cost of OpenAI/Google APIs.
Sources
- Black Forest Labs (April 27, 2026): Introducing FLUX Pro 2
- Artificial Analysis benchmark report (April 28, 2026)
- HuggingFace FLUX Pro 2 release page