Black Forest Labs publishes Self-Flow

Black Forest Labs has released Self-Flow, a self-supervised flow matching paradigm that outperforms REPA across audio, video, and image generation.

A team comprising Black Forest Labs (Patrick Esser, Dominik Lorenz, Robin Rombach) and academic researchers (Hila Chefer, Antonio Torralba, among others) publishes Self-Flow, a self-supervised flow matching paradigm that learns representation and generation jointly, without relying on an external encoder like DINOv2, SigLIP2, or MERT.

The core mechanism, dubbed Dual-Timestep Scheduling, applies heterogeneous noise levels to tokens of the same input. This informational asymmetry forces the model to reconstruct what is missing from a partially corrupted view, which leads to the emergence of a strong semantic representation in parallel with the generative capability. The approach thus dispenses with the alignment models usually grafted to stabilize the convergence of diffusion or flow matching models.

The results are evaluated across three modalities with a 4-billion-parameter FLUX.2 backbone jointly trained on 200M images, 6M videos, and 2M audio-video pairs, as well as on smaller models (625M) for isolated comparisons. Self-Flow outperforms REPA, the current benchmark for external encoder alignment, across all three modalities. The gap widens as the model scales up (from 290M to 1 billion parameters), whereas REPA's gains erode. Furthermore, for both audio and video, the authors note that external alignment struggles to generalize beyond image-centric tasks.

An avenue for future work is explored concerning world models: models trained with Self-Flow are fine-tuned for action prediction on the SIMPLER simulator, and maintain their advantage on complex multi-object or sequential tasks.