Chimera Hybrid DiT Reports 7.3x Compute Gain Over Wan-2.1

A new paper out of the Chimera group claims something that, if it holds up, matters more than the usual video-gen leaderboard skirmish. Their hybrid diffusion transformer reportedly hits 7.3x the compute efficiency of Wan-2.1's 2B baseline end-to-end, with 1.7x of that coming from the dense backbone alone before the sparsity tricks kick in. The paper on arXiv frames this as a Chinchilla-style scaling study for hybrid visual diffusion, which is the part worth pausing on.

The architecture itself is a stack rather than a single clean idea. Text, image, and video tokens are processed in one raster-ordered stream, and the model mixes Kimi Delta Attention, Multi-head Latent Attention, and sparse Mixture-of-Experts layers. Total parameter count is 11 billion, with 2 billion activated per token. The authors report that image pretraining wants compute divided nearly evenly between activated model size and training-token count, while video pretraining modestly favors model size at higher budgets. Those two ratios are the actual research contribution, more than any single benchmark number.

The headline demo is temporal extrapolation. Trained only on 5-second clips, the model is reported to generate coherent 30-second videos zero-shot, with the authors measuring only 6.5% FID degradation in the last five seconds. If that generalizes, the entire long-video training regime of paying for longer sequences at training time gets cheaper, because you can train short and stretch at inference.

The honest caveat is that all of this is one paper, comparing to one baseline the authors picked, on metrics they picked. FID is a weak proxy for whether a 30-second clip is actually watchable, and there is nothing in what I could pull about how the model compares to the current closed frontier of video generation or whether weights and training code will be released. Take the specifics as reported, not settled.

Still, the direction is the interesting part. If open-source video labs get a real compute-optimal recipe for hybrid DiTs instead of guessing scaling ratios, and if train-short-generate-long survives contact with production prompts, the economics of building a competitive video model shift toward smaller, faster-moving teams.