Saining Xie

Today, Dream Labs is coming out of stealth with Dream-1, a robot foundation model that controls bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour. Dream API is live for everyone, today. 🧵

🚀 What if video generators could build on representations that already understand the visual world? We are excited to introduce V-RAE: Rethinking Video Latent Spaces for Generation. Recent progress in image generation has begun to move beyond conventional VAE latents, exploring both direct pixel-space and representation-based approaches. Video generation, however, still depends heavily on latent compression, as the scale and redundancy of spatiotemporal data make direct modeling prohibitively expensive. However, most video VAEs are optimized for pixel reconstruction, and a latent space that reconstructs well is not necessarily easy to generate. V-RAE takes a different approach: it directly uses representations from frozen vision foundation models as the generative latent space, rather than as auxiliary supervision. We study DINOv3, SigLIP2, EUPE, and V-JEPA 2.1. A lightweight temporal attention pooling module compresses their dense features by 4×, followed by a spatiotemporal Transformer decoder. Under matched generation backbones, latent budgets, and training settings, V-RAE achieves: 🏆 2.13 rFVD on Kinetics-600 🎬 117.86 gFVD on UCF101 and 19.16 gFVD on Kinetics-600 ⚡ Up to 6× faster convergence than VAE-based latent spaces 🧠 90.92% semantic probing accuracy on UCF101 🌍 Better future prediction on Cityscapes, reducing gFVD from 144.47 to 111.36 Our experiments also reveal a broader finding: Good Reconstruction ≠ Good Generation. During generation, predicted latents inevitably deviate from real encoding trajectories. If the latent space is not sufficiently smooth, small errors can be amplified into visible artifacts. We therefore introduce tFVD to evaluate temporal smoothness and robustness to latent prediction errors. It correlates much more strongly with downstream generation quality, reaching 0.919 on Kinetics-600. The takeaway: A latent space is not merely where videos are compressed—it determines what the generator must learn. When semantics and temporal structure are already organized in the representation, generation becomes easier to learn. Representation first. Generation follows. Many thanks to my mentors, @ScottNLP and @SQWu_Tori, for their continuous guidance and support. I am also deeply grateful to @sainingxie for his valuable guidance and invaluable feedback, which greatly helped shape V-RAE. 🙏 Hi @_akhaliq, we would truly appreciate your help in sharing V-RAE with the broader AI research community. Thank you! 🙏 📄 Paper: arxiv.org/abs/2608.13556 💻 Code: github.com/V-RAE/V-RAE 🤗 Models: huggingface.co/Guomh0707/V-R… 🌐 Project: v-rae.github.io #VideoGeneration #GenerativeAI #ComputerVision #WorldModels #RepresentationLearning #RAE