Chimera Fits Chinchilla-Style Scaling Laws to a Hybrid Visual Diffusion Backbone, 7.3x More Efficient End-to-End

arXiv·medium signalXBlueskyLinkedInCopy link

Chimera processes text, image, and video tokens as one raster-ordered stream with no positional embeddings, combining Kimi Delta Attention for O(N) long-context state tracking, interleaved Multi-head Latent Attention for global interaction, modality-aware short convolutions for local spatiotemporal context, and sparse MoE layers. The contribution beyond the architecture is HeteroP, a module-wise hyperparameter transfer scheme keyed to each tensor's functional fan-in and model depth, which produces a consistently tuned family that can actually be fit to compute-optimal laws for activated size, token count, and image-video data ratio. The resulting 11B model with 2B activated parameters is 1.7x more compute-efficient than a full-attention dense backbone (7.3x on the complete system) and extrapolates zero-shot from 5-second training clips to 30-second video with 6.5% FID degradation in the final five seconds.