RoAR-VAE improves latent diffusability via robust prediction
Original title: RoAR-VAE: Improving Latent Diffusability through Robust Prediction
It beats the strongest baseline tokenizer by 2.19 points and FLUX.1-VAE small-text accuracy by 16.93 points.
RoAR-VAE is a visual tokenizer that keeps generation-relevant information recoverable from corrupted latents. RobustAlign predicts pretrained semantic features from corrupted latents, while RobustRecon reconstructs the clean image for spatial and appearance detail. The authors say robust prediction makes useful information more accessible in high-variance latent directions that diffusion models learn earlier. Under a matched training budget, average generation benchmarks rose 2.19 points over the strongest baseline tokenizer.
- Dropping REPA-E's diffusion objective still matched downstream generation under a matched setting.
- Small-text reconstruction accuracy rose 16.93 percentage points over FLUX.1-VAE.
- The gain is reported under the same training budget as the strongest baseline tokenizer.
From the original article
1 min read at huggingface.co
An effective visual tokenizer must preserve image fidelity while producing latents that downstream diffusion models can learn efficiently. Prior work improves latent diffusability through semantic alignment or joint tokenizer--diffusion training.
By Yang Yue, Fangyun Wei, Yitong Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.