Purrception: Variational Flow Matching for Vector-Quantized Image Generation
This work addresses image generation efficiency for machine learning researchers, presenting an incremental improvement by adapting existing flow matching techniques to vector-quantized latents.
The paper tackled the problem of vector-quantized image generation by introducing Purrception, a variational flow matching method that combines continuous transport dynamics with discrete supervision, resulting in faster training convergence on ImageNet-1k 256x256 and competitive FID scores compared to state-of-the-art models.
We introduce Purrception, a variational flow matching approach for vector-quantized image generation that provides explicit categorical supervision while maintaining continuous transport dynamics. Our method adapts Variational Flow Matching to vector-quantized latents by learning categorical posteriors over codebook indices while computing velocity fields in the continuous embedding space. This combines the geometric awareness of continuous methods with the discrete supervision of categorical approaches, enabling uncertainty quantification over plausible codes and temperature-controlled generation. We evaluate Purrception on ImageNet-1k 256x256 generation. Training converges faster than both continuous flow matching and discrete flow matching baselines while achieving competitive FID scores with state-of-the-art models. This demonstrates that Variational Flow Matching can effectively bridge continuous transport and discrete supervision for improved training efficiency in image generation.