Flux

Tested FLUX models over three weeks across the official fal.ai hosted endpoints plus a local install of FLUX.1 dev on a single RTX 4090. Used cases included generating brand creative for two product launches, rapid art-direction exploration, and a self-hosted comparison run against Leonardo and Midjourney on identical prompts. Baseline was Leonardo AI for SaaS image-gen and Stable Diffusion XL for the prior open-source workflow.

FLUX.1 pro produced output quality genuinely competitive with Midjourney v6 on roughly 60 percent of test prompts, particularly on photorealistic subjects and architectural scenes - the prompt adherence is visibly better than SDXL by a large margin. Text rendering inside images was the best I have seen from any open-source model, scrambling letters far less often than Leonardo or Midjourney on the same prompts. Self-hosting FLUX.1 dev on a 4090 actually worked at roughly 8 seconds per image at 1024 by 1024 - real production speed for a local workflow. The Apache-licensed weights for schnell variant make commercial use straightforward.

Hand and face quality on closeups still lags Midjourney v6 - had to regenerate roughly three out of ten portrait gens. The hosted pricing on fal.ai or Replicate for FLUX.1 pro runs higher per image than Leonardo Pro tier at scale, which surprised me - the open-source label does not translate to cheaper inference unless you self-host. No native batch UI - you assemble grid output yourself via ComfyUI nodes or API scripting, which raised the skill floor compared to Leonardo. Brand-consistency workflows like style references require LoRA training, not point-and-click. No mobile workflow at all.

Best for technical creatives, ML engineers, and brands with privacy requirements who can self-host. The quality bar for an open model is genuinely impressive. Midjourney v6 remains the right pick for non-technical creatives who want best-in-class single-shot quality. Leonardo AI wins for marketing teams who need brand-consistency batch workflows out of the box. Pick FLUX when ownership of weights and inference matters.