RoAR-VAE lifts generation 2.19 pts, small-text recon 16.93 pts

A new tokenizer paper on arXiv, RoAR-VAE, reports a 16.93 percentage-point jump in small-text reconstruction accuracy over FLUX.1-VAE, and a 2.19-point average lift on generation benchmarks over what the authors call 'the strongest baseline tokenizer' at a matched training budget.

The setup is a revisit of REPA-E. The authors write that 'removing its diffusion objective and dedicated generative blocks while retaining semantic prediction from corrupted latents yields comparable downstream generation performance under a matched setting.' From that they generalize to a principle they call robust prediction: 'generation-relevant information should remain recoverable from corrupted latents.' RoAR-VAE instantiates it with two objectives — RobustAlign, which 'predicts pretrained semantic features from corrupted latents to preserve high-level structure,' and RobustRecon, which reconstructs the clean image.

The abstract does not name which benchmarks make up the 2.19-point average, and it does not identify the 'strongest baseline tokenizer' being beaten. The FLUX.1-VAE comparison is restricted to the small-text reconstruction number.