Descriptions:
Stefan Batifol, developer relations engineer at Black Forest Labs (BFL), presents the company’s model roadmap and research direction at the AI Engineer conference, covering the full FLUX family from its August 2024 debut through to the company’s current work on what it calls visual intelligence. Flux 1 launched as an open-source text-to-image model and briefly became the most-liked model on Hugging Face; Flux Context followed as the first open-source model combining text-to-image generation with image editing in a single architecture, generating results in seven to eight seconds compared to the forty-plus seconds typical of contemporary competitors including the original GPT image generation. Flux 2, released in November, is the company’s current flagship, and Batifol presents photorealistic samples across portraits, animals, and product photography that he argues are indistinguishable from photographs.
The second half of the talk covers SelfFlow, a research paper BFL published approximately six weeks prior and made fully open. SelfFlow tackles a core limitation in current generative model training: the reliance on external encoders like DINOv2 that introduce scaling ceilings, modality specialization constraints, and objective misalignment — highlighted by a counterintuitive finding that DINOv3, technically superior to DINOv2, produces worse generative model performance when used as an alignment target. SelfFlow replaces external encoders with a self-supervised approach that combines representation learning and generation within the same flow, demonstrating 70x faster convergence during training. BFL’s current enterprise customer base includes Microsoft, Adobe, Canva, and Mistral.
📺 Source: AI Engineer · Published May 08, 2026
🏷️ Format: Keynote Launch
