Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis

Self-Flow, FLUX 3

prior work: REPA: Align diffusion models’ internal representations with DINO’s, since diffusion models don’t learn semantic representations well on their own.

  • con: this depends on an external model, scales poorly

Self-Flow: noise some tokens asymmetrically, align representations of the noisy tokens with the clean ones in addition to the generation objective

Making it multimodal:

  • per-modality in/out projection layers, autoencoder
    • FLUX.2 AE for images, WAN2.2 for video, Songbloom for audio
  • hparams like noise scheduling (same level of noise is harder to predict for images than video bc more info with video) / batch size
  • vector timestep conditioning