FLUX, Open Research, and the Future of Visual AI — Stephen Batifol, Black Forest Labs

Stephen Batifol, Developer Relations Engineer at Black Forest Labs (BFL), presents the company's trajectory from image generation toward a broader vision of visual intelligence — encompassing images, video, audio, actions, and physical AI. BFL is the team behind Stable Diffusion and Latent Diffusion, with over 200,000 academic citations and enterprise customers including Microsoft, Adobe, Canva, and Mistral.

The talk traces BFL's model releases chronologically. FLUX.1 (August 2024) was a text-to-image breakthrough that outcompeted Stable Diffusion, ran on consumer laptops, and became the most-liked model on Hugging Face at launch. FLUX Kontext followed as the world's first open-source combined text-to-image and image-editing model, generating and editing in 7–8 seconds versus GPT Image's 40–50 seconds at the time, and enabling storyboard-style character-consistent image sequences. FLUX.2 (November 2024) introduced multi-reference editing — accepting up to 10 simultaneous input images — delivering photorealistic outputs for product photography, outfit composition, and e-commerce visualization, achieving state-of-the-art results in both text-to-image and image editing in the open-source landscape. FLUX.2 Klein (January 2025) pushed further into real-time generation, achieving 300ms for generation and 500ms for editing, compared to Qwen's ~15–20 seconds on equivalent benchmarks while remaining at parity or better on quality metrics for 4B and 9B model sizes.

The central research contribution discussed is SelfFlow, a paper BFL released approximately six weeks before the talk. SelfFlow addresses a fundamental limitation of current generative model training: the reliance on external encoder models (such as DINOv2) for representation alignment. The standard approach adds noise to training images and denoises them, using an external encoder to teach the model physical and spatial understanding (e.g., that objects should not intersect). While effective — convergence is 70x faster with alignment — this introduces a scaling ceiling, modality specialization constraints, and objective misalignment. Notably, upgrading from DINOv2 to DINOv3 actually degraded generative model performance, illustrating the fragility of the dependency.

SelfFlow replaces external encoders with a self-supervised dual-model architecture: a student model receives high-noise inputs and learns both generation and representation jointly, while a teacher model (a stabilized exponential moving average of the student) receives low-noise inputs. The student minimizes both generation loss and representation loss simultaneously. Research models trained with SelfFlow outperformed flow-matching baselines across all three modalities — audio, images, and video — while converging faster and avoiding the plateau observed in baseline training. Practical improvements are visible in text rendering accuracy (correct letterforms within generated scenes), anatomical coherence in human figures, and temporal stability in video generation (eliminating flickering artifacts).

SelfFlow also enables joint multimodal generation from a single model — the same model trained on images can generate video, audio, and predict robotic actions. The talk shows a robot manipulation task where the SelfFlow model correctly picks up and moves an object while the baseline produces erratic arm movements, highlighting the approach's relevance to physical AI and robotics.

BFL's stated roadmap points toward world models — generative models that understand geometry, spatial relationships, and physical interactions — as the substrate for training autonomous agents, scaling self-driving systems, and automating manufacturing. The closing framing positions real-time interactive visual generation as the bridge between current image/video AI and future embodied AI.