Chimera

Chimera is a hybrid visual diffusion transformer co-designed with a principled, compute-optimal scaling recipe for the emerging “token-extensive” regime, where high-resolution images, long videos, and rich multimodal context make the quadratic cost of full-attention diffusion transformers increasingly prohibitive. Rather than treating model design and scaling as separate concerns, Chimera develops them jointly: the architecture defines a controlled family of models, and the scaling analysis decides how best to spend a fixed compute budget across that family.

Key ideas

• Design and scaling, co-developed. Architecture and scaling are studied together, so each design choice can be tuned once and extrapolated predictably rather than re-engineered at every scale.
• A unified multimodal stream. Text, images, and video are processed by one shared backbone, enabling direct cross-modal interaction without modality-specific branches.
• A hybrid, controlled model family. The backbone mixes complementary layer types — linear attention for efficient long-context state tracking, periodic global attention to restore token-to-token interaction, modality-aware convolutions for local spatiotemporal structure, and sparse mixture-of-experts for capacity at controlled activated compute.
• Minimal inductive bias, stable and deep. A simple raster scan with no positional embeddings, paired with stabilizing connections and normalization, keeps deep heterogeneous stacks robust to train.
• Compute-optimal scaling. A hyperparameter-transfer scheme lets a recipe tuned on a small proxy model transfer reliably across widths and depths, making it possible to fit Chinchilla-style laws over model size, training tokens, and image–video data composition.

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 adapts these architectural and scaling lessons to the distinct structure of visual data, offering a principled foundation for efficient long-context generative backbones whose behavior can be reasoned about ahead of expensive large-scale training.

This work is currently under review; more details will follow.