Detect Flux AI Images

What Flux.1 is, and why it is hard

Flux.1 is a family of AI image generators developed by Black Forest Labs, a company founded by former Stability AI researchers including Robin Rombach, the lead author of the original Stable Diffusion paper. Released in 2024, Flux.1 achieved state-of-the-art image quality that surpassed Midjourney v6 and SDXL on multiple benchmark evaluations: particularly for photorealistic portraits, coherent text rendering, and complex compositional scenes. Flux.1 is architecturally distinct from previous diffusion models. Traditional diffusion models like Stable Diffusion and Midjourney use score-based stochastic differential equations to progressively denoise images. Flux uses rectified flow matching, a different mathematical framework that defines straighter trajectories from noise to image in the sampling process. This produces cleaner, less noisy images with fewer of the visible artifacts that made earlier SD outputs easier to detect. The dual-stream transformer architecture (DiT - Diffusion Transformer) processes image patches and text conditioning in parallel streams before merging, which produces a different pattern of spatial attention artifacts than the U-Net architectures used by SD 1.5 and SDXL. Flux.1 Dev is open-source with 12 billion parameters - significantly larger than SDXL's 6.6B, and runs locally without any API or moderation layer, making it the preferred tool for high-quality fraud imagery in 2025 and 2026.

The fingerprints, and how they differ across Dev, Pro and Schnell

Flux images are cleaner than most diffusion output, and that cleanliness is itself a signature. The rectified flow matching sampling process produces a different noise profile than score-based diffusion. Specifically, the flow matching trajectories converge more directly from noise to image, which means the intermediate noise states are different, and the residual noise characteristics in the final image reflect this different sampling path. In the frequency domain, Flux images show characteristic spectral signatures at spatial frequencies that differ from both Stable Diffusion and Midjourney. The dual-stream transformer architecture produces distinctive attention artifacts, subtle patterns in regions where the text-conditioned and image-conditioned attention streams merge, that are detectable in the pixel distribution even at full resolution. PRNU (Photo Response Non-Uniformity) is completely absent in Flux outputs, as in all AI generators, because no camera sensor was involved. The combination of flow-matching noise characteristics, dual-stream attention artifacts, and PRNU absence provides FauxLens with multiple independent signals for Flux detection that persist through typical post-processing including JPEG compression and moderate resizing.

The three main Flux.1 variants differ in sampling steps, architecture details, and access method, which produces forensically distinguishable outputs. Flux.1 Dev is the open-source, full-quality variant with 12B parameters. It uses the full rectified flow sampling schedule (typically 20-50 steps) and produces the highest quality outputs with the most well-defined frequency-domain signature. Detection accuracy on Flux Dev outputs exceeds 91%. Flux.1 Schnell is a distilled variant trained to produce acceptable quality in 1-4 sampling steps - significantly fewer than Dev. The distillation process compresses the sampling trajectory, which changes the noise characteristics of the output. Schnell images show more compression-like artifacts than Dev outputs because the distilled model has fewer steps to refine detail. These artifacts make Schnell images slightly easier to detect than Dev outputs, despite Schnell's lower overall image quality. Flux.1 Pro is an API-only commercial variant that Black Forest Labs periodically updates with undisclosed architectural refinements. Pro outputs show slightly different frequency signatures than Dev outputs, reflecting these optimizations. FauxLens maintains detection models for all three variants and is updated when Black Forest Labs releases new Pro versions. Community fine-tunes based on Flux Dev (available on Civitai and HuggingFace) inherit the Dev forensic fingerprint while adding style-specific artifacts.

Where Flux images have shown up

Flux.1's combination of open-source availability, high image quality, and local execution without moderation has made it the fastest-growing tool for AI-generated fraud imagery since its 2024 release. In romance scams, Flux has largely replaced Midjourney for operators who want consistent high-quality persona images without Discord attribution - Flux runs locally with no service trail. The FBI's Internet Crime Complaint Center noted a shift in the technical sophistication of romance scam imagery beginning in late 2024, coinciding with Flux.1's public release. In synthetic identity fraud, where AI-generated faces are used to create fraudulent identity documents, Flux's photorealistic portrait quality has made it the preferred tool. The document fraud pipeline typically involves generating a Flux portrait, then compositing it into a genuine document template. In political disinformation, Flux-generated imagery appeared in documented influence operations targeting the 2025 UK local elections and several European municipal elections, where photorealistic fabricated imagery of politicians was distributed through Telegram channels before fact-checkers identified it as AI-generated. Flux's lack of any visible watermark or platform attribution, unlike Midjourney's subtle Discord-linked metadata, makes attribution of Flux-generated disinformation significantly harder without forensic analysis tools.

Keeping up as the models move

Keeping Flux detection current is a continuous engineering challenge because Black Forest Labs actively develops the model and the community produces new fine-tunes weekly. FauxLens addresses this through three mechanisms. First, architecture-level detection: rather than training exclusively on specific Flux model outputs, our detection models learn the underlying rectified flow and dual-stream transformer artifacts that all Flux variants share. These architectural fingerprints persist across fine-tunes and minor version updates. Second, continuous retraining: as new Flux variants and fine-tunes become available, we incorporate their outputs into our training dataset and update detection models on a regular release cycle. Third, ensemble detection: our Bayesian evidence fusion combines six independent signals, which means that even when a new Flux variant shifts one signal's characteristics, the remaining signals continue to provide detection capability. This ensemble approach makes FauxLens resilient to the cat-and-mouse dynamic where new generator versions are specifically designed to evade detectors trained on older outputs. When Black Forest Labs releases major new Flux versions - such as the Flux 1.1 Pro update in late 2024 - we announce updated detection support in our changelog.