🎯 Prior normalizing flows (NFs) such as STARFlow add random noise to VAE latents as augmentation, leading to complex pipelines with extra noising and denoising steps. End-to-end training is appealing, but previous attempts (e.g. REPA-E) observed latent collapse when naively training VAEs and generative models together. 💡 We propose SimFlow — a simple, end-to-end training framework for NFs. The key idea is surprisingly simple: fix the variance predicted by the VAE encoder to a constant (e.g. 0.5^2). - Simple. The encoder outputs a broader latent distribution, and the decoder learns to reconstruct clean images directly — no extra noise or denoising design (unlike STARFlow). - End-to-end. Fixed variance simplifies the ELBO and stabilizes joint training of NF VAE. - Effective. SimFlow achieves new state-of-the-art among NFs on ImageNet 256×256 and 512×512. 🙌 Joint work with Guangting Zheng, Tao Yang, Rui Zhu, Xingjian Leng (@xingjian_leng), Stephen Gould (@sgould_au), and Liang Zheng (@LiangZheng_06). 🙏 Huge thanks to other colleagues at ByteDance and Hanhong Zhao for many inspiring discussions. 🔗 Project: qinyu-allen-zhao.github.io/S… ⭐ Code: github.com/ByteDance-Seed/Si… 🤖 Models: huggingface.co/QinyuZhao1116… 📄 Preprint: arxiv.org/abs/2512.04084