Flux

Flux is an open-source image generation model from Black Forest Labs, founded by the core researchers behind Stable Diffusion. Released in August 2024, it quickly became the new benchmark for open-weights image generation, with photorealism and prompt adherence that competes seriously with closed commercial models.

When Black Forest Labs released Flux in August 2024, the reaction from the AI image generation community was immediate and unusually uniform: this is better than we expected. Not a marginal improvement over existing open-source options, but a step change. The team that built this had previously created Stable Diffusion, and with Flux they applied what they'd learned from watching three years of community use, research iteration, and competitor models to build something that addressed most of the complaints about what came before.

I've been running Flux in various configurations since shortly after launch, and the honest assessment is that it lives up to the initial reaction. This review covers what Flux is, how the different model tiers work, where it competes with commercial tools, and who should actually be using it.

Quick verdict

Flux.1 [pro] via API produces images that are competitive with Midjourney on photorealism and significantly better than Stable Diffusion on prompt accuracy. Flux.1 [dev] as a free, locally-runnable model has no real equivalent in quality at its price point (zero). If you want the best open-source image generation available in 2026, Flux is the answer. If you want the most aesthetically polished images with the least effort, Midjourney is still ahead on stylized and artistic output.

What Flux is and who built it

Black Forest Labs is a research company founded in 2024 in Freiburg, Germany. The founding team includes Robin Rombach, Andreas Blattmann, and Patrick Esser, who were among the core researchers behind Stable Diffusion at Stability AI and CompVis at the University of Freiburg. When they left Stability AI and formed Black Forest Labs, they took with them the accumulated knowledge of building and iterating on what had become the most widely used open-source image model in the world.

The Flux architecture is based on flow matching rather than the denoising diffusion process used by Stable Diffusion. Flow matching is a different approach to learning the mapping between noise and images, and it has practical advantages in training efficiency and sample quality. The technical paper is available, but the practical result is that Flux produces better images with better prompt fidelity than SDXL, the previous state of the art among open-source models.

The name "Flux" refers to the model family. Within that family, there are three main variants, and understanding what each is for matters for making the right choice.

The three Flux.1 models explained

Flux.1 [pro] is the highest-quality model in the lineup. It's not publicly available for download. You access it via the Black Forest Labs API, Replicate, Together AI, Fal.ai, or a growing list of platforms that have licensed it. At roughly $0.05 per image, it's affordable for most production use cases. The output quality is competitive with Midjourney on photorealistic subjects and outperforms most alternatives on complex compositional prompts.

Flux.1 [dev] is the open-weights version of the high-quality model, released for non-commercial use. You can download the weights from Hugging Face and run them locally or fine-tune them on your own data. The quality gap between [dev] and [pro] is real but smaller than you'd expect given that one costs money and the other doesn't. For most creative or research use cases where commercial licensing isn't required, [dev] is the right starting point.

Flux.1 [schnell] is the speed-optimized variant, released under the Apache 2.0 license, which means fully commercial use. The tradeoff is quality: [schnell] is noticeably weaker than [dev] and [pro]. But it's extremely fast, and for applications where generation latency matters more than peak quality, [schnell] makes sense. It's also the version you'd use for rapid iteration workflows where you're testing many variations before doing final high-quality generation.

Image quality: the actual comparison

The most meaningful test is how Flux.1 [pro] compares to Midjourney v6 and DALL-E 3 on different prompt types.

On photorealism, Flux.1 [pro] is the strongest of the three. Portraits, product shots, architectural renders, and natural scenes all come out with a level of detail and technical accuracy that Midjourney doesn't consistently match. The textures on fabric, the light behavior on skin, the rendering of complex reflective surfaces: Flux handles these more correctly.

On artistic and stylized output, Midjourney is still ahead. Flux follows prompts for artistic styles accurately, but the aesthetic sensibility it brings to the output is more neutral. Midjourney has a point of view. The images it produces feel designed in a way that Flux's don't by default. For concept art, mood boards, and images where "looks like art" is the goal, Midjourney remains the benchmark.

On prompt adherence, Flux and DALL-E 3 are both ahead of Midjourney. Complex multi-part prompts describing specific subjects, relationships, settings, and moods are represented more accurately in the output. This is where the flow-matching architecture shows practical benefits.

Running Flux locally

Flux.1 [dev] runs on NVIDIA GPUs with 12GB VRAM comfortably. 8GB is possible with memory optimizations (using FP8 quantization via the Diffusers library, for example) but generation times increase. For Apple Silicon, 16GB unified memory runs [dev] reasonably well, and 24GB+ runs it fast. The M4 Max and M4 Ultra Macs that shipped in early 2025 are genuinely good Flux machines.

The practical setup involves downloading the weights from Hugging Face (around 24GB for the full [dev] model in BF16), installing the Diffusers library, and writing a generation script or using ComfyUI with Flux-specific custom nodes. It's similar in complexity to setting up SDXL with Automatic1111, which is to say: approachable if you're comfortable with Python environments, confusing if you're not.

ComfyUI has become the preferred workflow tool for Flux among serious users. The node-based interface lets you build multi-step pipelines including LoRA loading, upscaling, and conditioning, and the community has developed a rich set of Flux-specific nodes. If you're coming from a Stable Diffusion ComfyUI workflow, the transition is straightforward.

Fine-tuning and the growing ecosystem

One of the things that made Stable Diffusion so powerful was the enormous community of fine-tuned models and LoRAs available through CivitAI and Hugging Face. Flux is building a similar ecosystem, though it's still younger.

LoRA training for Flux works via the same general approach as Stable Diffusion LoRAs, but the training scripts need to be Flux-compatible. Tools like the Kohya scripts and ostris's AI Toolkit have Flux support. Given that [dev] is the base model for fine-tuning (not [pro], which isn't publicly available), the community fine-tunes are for [dev]. Quality varies significantly, as it does in any community model ecosystem.

By mid-2026, the Flux ecosystem on Hugging Face and CivitAI is substantial but still smaller than the Stable Diffusion ecosystem. If you need a specific artistic style or domain model, there's a reasonable chance something exists for Flux, but the Stable Diffusion catalog is still larger. This gap will close.

API pricing and hosted access

Flux.1 [pro] via the Black Forest Labs API is priced at approximately $0.05 per image for standard resolution. This is competitive with DALL-E 3 at $0.04-0.08 per image and cheaper than some premium tiers on platforms like Midjourney at scale.

Replicate and Together AI offer Flux.1 [pro] at similar per-image rates with their own API infrastructure. Both have strong reliability and fast inference, and their SDKs are well-documented. For applications where you want a simple API integration without managing infrastructure, either platform works.

The lack of an official Black Forest Labs web app for consumers is a notable gap. If you want to try Flux.1 [pro] without touching an API, you'll need to use a third-party platform. Several consumer-facing image generation tools, including NightCafe, Fal.ai, and others, have added Flux as a model option with their own UIs. The experience varies by platform.

Who should use Flux

Developers building image generation into applications who want API access to a high-quality open-source model. The per-image pricing is fair, the output quality is competitive, and the open-source nature means you have options if you eventually want to host the model yourself.

Researchers and practitioners who were previously on Stable Diffusion and want to move to a more capable base model. The technical similarity of the ecosystems (ComfyUI compatibility, LoRA support, Hugging Face distribution) makes the transition relatively smooth.

Privacy-focused users who need local image generation with strong output quality. Flux.1 [dev] locally is the best option in this category.

Studios and production teams who need photorealistic images at scale and want more control than Midjourney's subscription model provides. The API-based pricing and ability to eventually self-host the model gives more flexibility for large-scale workflows.

Casual users who want a simple, beautiful web app with minimal setup: this is not the right first stop. Try Midjourney or Ideogram first.

Flux vs the alternatives directly

Flux vs Stable Diffusion: Flux wins on out-of-the-box quality by a significant margin. Stable Diffusion wins on ecosystem maturity, variety of community models, and tooling breadth. If you're starting fresh today, start with Flux. If you have years of Stable Diffusion workflows and fine-tunes, evaluate whether the quality gain justifies migration.

Flux vs Midjourney: Flux wins on photorealism and prompt accuracy. Midjourney wins on stylized artistic output and ease of use. These are genuinely different strengths. Many serious image generation practitioners use both.

Flux vs DALL-E 3: Similar on prompt accuracy, Flux ahead on photorealism, DALL-E 3 ahead on ChatGPT integration and text rendering. If your workflow is API-based, Flux is competitive. If your workflow is ChatGPT-centered, DALL-E 3 is more integrated.

The honest take

Flux is the most important open-source image generation development since Stable Diffusion's original release. The founding team brought serious research credibility and they executed well on delivery. The model quality is real, the open-weights commitment is real, and the ecosystem is growing at a pace that suggests it will keep improving.

The gap relative to Midjourney on stylistic output is genuine but narrowing. In a year from now, I expect the comparison to be closer. In the meantime, Flux is the clear answer for anyone who values open-source flexibility, photorealism, or API-based generation. And for anyone who needs the best free local image generation available, Flux.1 [dev] is where you should be.

Key features

• Flux.1 [pro] model competitive with top commercial image generators
• Flux.1 [dev] open-weights model for local and fine-tuned use
• Flux.1 [schnell] optimized for fast inference at lower quality
• Strong photorealism and prompt adherence
• Flow-matching architecture for improved training efficiency
• LoRA fine-tuning support via community tools
• API access via Replicate, Together AI, and Black Forest Labs directly

Pros and cons

Who is Flux for?

• High-quality photorealistic image generation via API
• Local image generation for privacy-sensitive workflows
• Fine-tuning on custom datasets for branded image styles
• Developer applications requiring open-weights model access
• Production workflows needing strong prompt-to-image accuracy

Alternatives to Flux

If Flux isn't quite the right fit, the closest alternatives are stable-diffusion , midjourney , dall-e , and ideogram . See our full Flux alternatives page for side-by-side comparisons.