Back to Projects

Chimera

Designing and Chinchilla-scaling hybrid visual diffusion transformers

Chimera is a hybrid visual diffusion backbone with a principled, compute-optimal scaling recipe for the regime where high-resolution images, long videos, and multimodal context make the quadratic cost of full attention prohibitive. It processes text, image, and video tokens in a single raster-ordered stream with no positional embeddings, and it is co-designed with a scaling methodology so architecture and scaling are developed together rather than as separate concerns.

Chimera couples architecture design with compute-optimal scaling for hybrid visual diffusion.

Key ideas

  • A unified, position-embedding-free stream. Text, image, and video tokens are processed in one raster-ordered sequence without positional embeddings, enabling direct cross-modal interaction in a single backbone.
  • A hybrid attention design. Chimera combines Kimi Delta Attention (KDA) for long-context state tracking at O(N) complexity, interleaved Multi-head Latent Attention (MLA) for direct global interaction, and modality-aware short convolutions for local spatiotemporal context.
  • Sparse mixture-of-experts. MoE layers expand capacity while keeping the activated compute controlled.
  • HeteroP scaling. A module-wise scheme transfers hyperparameters across width and depth according to each tensor’s functional fan-in and model depth, yielding a consistently tuned family for fitting Chinchilla-style compute-optimal laws over activated model size, training tokens, and image–video data ratio.

Results

  • Compute efficiency. Measured by pretraining diffusion loss, the dense backbone is 1.7× as compute-efficient as a matched full-attention Wan-2.1 2B baseline; the complete system reaches 7.3×.
  • Zero-shot length extrapolation. Without length-specific fine-tuning, Chimera extrapolates from 5-second training clips to 30-second videos with only 6.5% FID degradation in the last five seconds.
  • Compute-optimal recipe. Guided by the fitted laws, the team trains an 11B-parameter Chimera with 2B activated parameters; image pretraining splits compute nearly evenly between model size and tokens, while video pretraining modestly favors model size at higher budgets.

Scaling laws

Left: at a fixed training loss, the full Chimera system reaches the same quality with up to 7.3× less compute than a matched full-attention Wan-2.1 baseline. Right: fitted Chinchilla-style loss–compute frontiers across model sizes, used to pick compute-optimal configurations.

Video demo

Zero-shot 30-second generation (16 fps, 481 frames) from the same prompt, compared across Wan2.1, HunyuanVideo, and Chimera. “motion” is the mean frame-to-frame pixel change — higher means more actual movement, while a value near zero is a near-frozen clip.

A ceramic bowl tips over on a smooth kitchen countertop, spilling rice grains that scatter and pile into a small mound.

Wan2.1 motion 1.62
HunyuanVideo motion 0.76
Chimera ours motion 3.39

A single yellow rose in full bloom is featured in a tranquil garden while the camera pans right as the petals slowly open.

Wan2.1 motion 0.84
HunyuanVideo motion 0.83
Chimera ours motion 12.09

Why it matters

Visual generation is following language models into a long-sequence regime, but the language recipe does not transfer directly — visual diffusion must preserve spatiotemporal locality and bidirectional cross-modal interaction. Chimera establishes a foundation for designing and scaling efficient long-context diffusion architectures whose behavior can be reasoned about ahead of expensive large-scale training.

Read the paper: Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers.

Back to Projects