Jointly training the VAE and the diffusion model from scratch collapsed the latent space until a pixel-reconstruction loss was added, then cut training time by over 140x against vanilla SiT
AspireAI community · arXiv (cs.CV) - CVPR 2026
The usual recipe for latent image generation trains a VAE first and freezes it, then trains a diffusion model on top. SpeeDiff tries the joint version - training both from scratch so the diffusion gradient can shape the latent space into something generation-friendly - and the interesting part is the failure it ran into. Naive end-to-end backpropagation of the diffusion loss collapses the latent space outright. The fix is a Tweedie Pixel Reconstruction loss: decode a predicted clean latent from an intermediate noisy state using Tweedie's formula, which gives the encoder pixel-level feedback and keeps the latent space from collapsing.
With that loss in place, the paper reports training that is over 140x faster than Vanilla SiT and 61x faster than REPA, while reaching an FID of 1.50 without guidance on ImageNet 256x256. A more efficient 32x-compressed VAE variant reaches FID 1.53 without guidance on ImageNet 512x512. The architecture is fully transformer-based and is scaled jointly rather than staged.
The honest caveat is that both speed multipliers are relative to specific prior setups in the authors' own runs, so they describe a like-for-like comparison inside one paper rather than a universal speedup. The quality evidence is automated FID on ImageNet at two resolutions with no human evaluation, and the paper makes no claim that the joint-VAE approach transfers to video or text.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.