Flux

The Team That Built Stable Diffusion Started Over

Black Forest Labs launched in 2024 with a stacked roster: the original creators of Stable Diffusion and Latent Diffusion Models. They raised over $100 million before shipping a single product. Then Flux dropped, and it was immediately clear why investors wrote those checks. The image quality from Flux.1 Pro competes head-to-head with Midjourney V6 and DALL-E 3, particularly on photorealism and prompt adherence.

Three Models, Three Use Cases

Flux comes in three variants. Schnell is the speed demon — it generates images in 1-4 steps, under a second on good hardware. Quality takes a hit, but for prototyping or bulk generation, the speed is unmatched. Dev sits in the middle: open-weight, good quality, 20-30 steps. Pro is the flagship — closed-weight but available through the API at roughly $0.04-$0.06 per image depending on resolution.

Most developers start with Dev for testing, then switch to Pro for production. The quality difference between Dev and Pro is noticeable but not dramatic. Schnell is a different beast entirely — fast and rough, useful when you need thousands of images and don't need perfection.

Self-Hosting Changes the Economics

Here's where Flux gets interesting for businesses. Dev and Schnell are open-weight models you can run on your own infrastructure. Got an A100 GPU sitting in a server rack? Run Flux locally with zero per-image costs. Companies generating 50,000+ images monthly save thousands compared to API pricing. The model weights are available on Hugging Face with a permissive license for Dev.

RunPod, Replicate, and Together AI all offer hosted Flux endpoints if you want the API convenience without managing GPUs. Prices typically run $0.01-$0.04 per image through these platforms, undercutting Black Forest Labs' own API.

What Flux Does Better Than Everyone Else

Text rendering. Flux handles text in images with accuracy that rivals DALL-E 3 and crushes Midjourney. Need a poster with a specific headline? Flux nails it on the first try more often than not. Hands are also notably better — fewer extra fingers, more natural poses. Prompt adherence is exceptional; the model does what you ask without creative reinterpretation.

The Rough Edges

Documentation is sparse. The API reference exists but lacks examples for edge cases. Community resources fill the gaps, but you'll spend time on Reddit and Discord figuring things out. The ecosystem is young — no official plugins for Photoshop or Figma yet, though ComfyUI support is solid.

No built-in image editing. Unlike DALL-E with its inpainting or Midjourney with vary/zoom, Flux generates from scratch. You need separate tools or custom pipelines for iterative editing. ControlNet support exists through the community but requires technical setup.

Who Should Pick Flux

Developers and technical teams who want the best price-to-quality ratio in image generation. If you can handle a bit of infrastructure, Flux gives you Midjourney-level quality at a fraction of the cost. Solo creators without technical skills should look elsewhere — the learning curve is real.