Flux by Black Forest Labs

Flux is the open-weights image generation family from Black Forest Labs, founded by ex-Stability AI researchers. The headline model, FLUX.1, hit the scene in 2024 and immediately challenged Midjourney on prompt fidelity.

There are three Flux variants. Pro is the high-end commercial API. Dev is open weights for non-commercial use. Schnell is open weights, Apache-licensed, and fast.

If you've seen viral text-to-image gallery posts in the last year, half of them were Flux output. The model rendered hands and text better than anything else open at the time.

What Flux actually does

Flux generates images from text prompts. It also handles image-to-image variation, inpainting, and (via the Pro API) Flux Tools features like fill, redux, and depth control.

The architecture is a hybrid transformer plus diffusion approach with rectified flow training. That's why prompts hit harder. Token-to-pixel alignment is unusually good.

You can run Flux Schnell and Dev locally on a 24GB GPU. Flux Pro is API-only and runs on Black Forest Labs' infrastructure or via partners like Replicate, Fal, and Together.

Who Flux is for

Indie creators who want Midjourney quality without the Discord workflow love Flux. The Schnell variant runs locally on a beefy laptop or a single 4090.

Product teams building AI features (avatar generators, marketing image tools) plug Flux Pro into their stack via API. License terms favor commercial deployment more than older Stability releases.

Researchers and ComfyUI tinkerers love that weights are public. The community has trained countless LoRAs and fine-tunes on top of Flux Dev.

Pricing breakdown

Flux Schnell is free, Apache 2.0, and commercially usable. Flux Dev is non-commercial only via the BFL license. Flux Pro is API-only with usage-based pricing.

Replicate prices Flux Pro at around $0.055 per image. Fal and Together offer similar rates with their own tweaks.

Self-hosted Flux Schnell costs you GPU time, not API tokens. That's a big deal for high-volume use cases.

Standout features

Prompt fidelity is the big win. Flux follows long, detailed prompts better than most open models. Hands and text rendering are notably stronger.

Flux Pro's variants (Pro 1.1, Ultra, Raw) push fidelity, resolution, and style further. Flux Tools include fill (smart inpainting), redux (style transfer), and canny/depth control.

Flux Kontext, the newer release, layers character consistency and editing controls on top. It's how you keep a character looking the same across a 20-image series.

Honest tradeoffs

Open weights for Schnell are great. Flux Dev is non-commercial only, which trips up startups who don't read the license carefully.

Out of the box, Flux is heavier than SDXL. Inference takes longer per image on consumer GPUs. Optimized loaders and quantization help, but it's still a memory-hungry model.

The BFL ecosystem is young. Documentation is improving but lighter than Stability's. You'll lean on Replicate or Fal for production-grade serving.

Flux vs Midjourney vs Stable Diffusion

Midjourney has the cleanest aesthetic but no API and no weights. Stable Diffusion is open and has the deepest community ecosystem. Flux sits between them: open Schnell weights, Pro-quality API, and prompt fidelity that beats both on certain prompts.

See best AI image generators and Midjourney alternatives. Our Flux vs Midjourney comparison goes deeper.

Other Flux alternatives: Ideogram (great at text), Recraft, and DALL-E 3. The space moves monthly, so check the comparison pages for current standings.

Bottom line on Flux

Flux pushed open-weight image generation forward in a way nothing has since SDXL. It's the model to fine-tune on if you're shipping AI image features in 2026.

For more, see tools for AI creators and the Midjourney profile. Flux changed the calculus on whether you can self-host generative image models. The answer is now clearly yes.

Try Schnell locally first. Move to Pro if you need the highest fidelity.

Self-hosting Flux Schnell on consumer GPUs

Flux Schnell runs on a single 24GB GPU like the RTX 4090 or 3090. With quantization tricks (FP8, NF4), you can squeeze it onto cards with less VRAM.

ComfyUI is the most common runtime in the community. It exposes Flux's nodes and lets you build complex pipelines with LoRAs, ControlNet, and image-to-image flows.

For production, dedicated inference servers like Replicate or Fal handle scaling, queueing, and reliability. Self-hosted is great for experimentation. Hosted is better for product features.

LoRAs and fine-tunes for Flux

The community has trained countless LoRAs on top of Flux Dev. Style LoRAs (anime, ink wash, photorealistic), character LoRAs, and concept LoRAs all exist.

Civitai is the main sharing hub. Quality varies. Trained character LoRAs let you keep a consistent character across many image generations, which is huge for storytelling and ad creative.

Fine-tuning Flux Dev requires meaningful compute (a single A100 or rented multi-GPU). Tools like Replicate's training endpoints make it easier without your own hardware.

Flux Kontext and character consistency

Flux Kontext is the newer Black Forest Labs release focused on consistent characters and editing. Drop a reference image and a prompt, get a controlled variation.

This solves one of the longest-standing pain points in generative image: keeping a character looking the same across a 20-image sequence. Kontext doesn't fully solve it, but it pushes the bar far forward.

For comic creators, ad agencies, and storyboard artists, Kontext is the most consequential image-AI release of the year.

Final word on Flux

Flux changed the open-weights conversation in 2024 and keeps shipping in 2026. The team's research pedigree is the moat. Expect more releases.

For commercial AI image features, Flux Pro or Flux Schnell (Apache-licensed) are both viable. Pick based on the quality-vs-cost tradeoff your product needs.

Flux Pro for production AI image features

Flux Pro is the API-only tier from Black Forest Labs. It runs on BFL's infrastructure or via partners like Replicate, Fal, Together, and Runware.

For product teams building avatar generators, ad creative tools, or AI photo apps, Flux Pro hits the quality bar that consumers expect. Output rivals Midjourney on many prompts.

Flux Pro 1.1 (sometimes called Pro Plus) and Flux Ultra extend resolution, fidelity, and edge-case quality. Flux Raw produces less stylized output for editorial use.

Pricing varies per partner. Replicate, Fal, and Together publish per-image rates. For high-volume use, negotiate directly with BFL for enterprise pricing.

Latency is fast for a frontier image model. Most providers return generations in under 10 seconds. Streaming is supported via partners that offer it.

For commercial use, double-check the license tier. Flux Pro and Flux Schnell are commercially safe. Flux Dev is non-commercial only and trips up startups who skim the license.

For more AI image options, see best AI image generators and our Midjourney, DALL-E 3, and Ideogram profiles.

Flux FAQ

Is Flux better than Midjourney? Different strengths. Flux follows complex prompts more literally and renders text and hands more reliably. Midjourney has a stronger native aesthetic and tighter style consistency. Try both for your specific use case.

Can I run Flux on a Mac? With limitations. Flux Schnell can run on M-series Macs via tools like DiffusionBee or ComfyUI with MLX backends, but performance lags behind a real GPU. For serious local work, an NVIDIA card wins.

What's the license situation? Flux Schnell is Apache 2.0 (commercially safe). Flux Dev is non-commercial only. Flux Pro is commercial via API. Read the license carefully before shipping a product on Flux Dev.

How does Flux compare to Stable Diffusion 3? Flux models tend to render more cleanly out of the box. SD3 has a deeper community ecosystem with more LoRAs and fine-tunes. Both are valid choices depending on your needs.

Does Flux support image-to-image and inpainting? Yes. Both are native capabilities. Flux Tools (via the Pro API) extend this with smart fill, redux style transfer, and depth or canny conditioning.

For developers shipping AI image features, Flux deserves a serious evaluation. The combination of open weights for indie experimentation and Pro API for production scale is a rare alignment in the AI tooling landscape.

Flux community and ecosystem

The Flux community on Hugging Face, Civitai, and Reddit is active. LoRAs, fine-tunes, training scripts, and ComfyUI workflows are shared daily.

For practitioners, this ecosystem is a force multiplier. Need a Studio Ghibli-style LoRA? Someone has trained one. Need a specific anime aesthetic? Someone has uploaded it.

Black Forest Labs maintains an active research blog. New releases (Flux Pro 1.1, Ultra, Raw, Kontext) are documented with technical details and example outputs.

For commercial use, the licensing structure rewards careful reading. Flux Schnell (Apache) and Flux Pro (commercial) are safe. Flux Dev is non-commercial only and trips up startups who skim.

The training scripts for fine-tuning Flux Dev on consumer hardware exist but require meaningful technical comfort. Renting H100 time on Replicate or Runpod is often more practical than local fine-tuning.

For more AI image options, see our best AI image generators roundup and Midjourney, Stable Diffusion, and Ideogram profiles.

Flux pushed open-weight image generation forward in 2024 and continues setting the bar in 2026. For developers shipping AI image features, it deserves serious evaluation.