Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

• URL: http://arxiv.org/abs/2607.28611v1
• Date: Thu, 30 Jul 2026 17:58:02 GMT
• Title: Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers
• Abstract summary: Visual generation requires high-resolution images, long videos, and multimodal context.<n>We introduce Chimera, a hybrid visual diffusion backbone with a principled scaling recipe.<n>We show that the dense backbone is 1.7x as compute-efficient as a matched full-attention Wan2.1 2B baseline.
• Score: 54.01042826416009
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Visual generation increasingly requires high-resolution images, long videos, and multimodal context, making the quadratic cost of full attention prohibitive. We introduce Chimera, a hybrid visual diffusion backbone with a principled scaling recipe. Chimera processes text, image, and video tokens in one raster-ordered stream without positional embeddings. It combines Kimi Delta Attention (KDA) for long-context state tracking with 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 controlling activated compute. To scale this heterogeneous architecture, we introduce HeteroP, a module-wise scheme that transfers hyperparameters across width and depth according to each tensor's functional fan-in and model depth. HeteroP yields a consistently tuned family used to fit Chinchilla-style compute-optimal laws for activated model size, training-token count, and image-video data ratio. Guided by these laws, we train an 11B-parameter Chimera with 2B activated parameters. Experiments show three results. First, measured by pretraining diffusion loss, the dense backbone is 1.7x as compute-efficient as a matched full-attention Wan-2.1 2B baseline, while the complete system reaches 7.3x. Second, without length-specific fine-tuning, Chimera extrapolates zero-shot from 5-second training clips to 30-second videos, with only 6.5% FID degradation in the last five seconds. Third, the fitted laws show that compute-optimal image pretraining divides compute nearly evenly between activated model size and training-token count, whereas video pretraining modestly favors model size at higher budgets. These results establish a foundation for designing and scaling efficient long-context diffusion architectures.

Related papers

• LLaDA MoE v2: Scaling Mixture-of-Experts Diffusion Language Models [129.64558694024126]
  Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling.<n>We characterize how optimization hyper parameters, compute allocation, and architecture scale for MoE dLLMs.<n>We train LLaDA MoE v2, a 30B-A3B dLLM, from scratch on 23.5T tokens.
  arXiv  Detail & Related papers  (2026-08-04T10:53:02Z)
• DiffusionGemma Technical Report [72.10570774996219]
  We introduce DiffusionGemma, a language model that uses discrete diffusion to generate text at exceptionally high speed.<n>Rather than decoding one token at a time, DiffusionGemma iteratively refines blocks of 256 tokens in parallel.<n>Our compute-efficient two-stage training pipeline uses fewer than 10% of the starting AR model's total training token budget.
  arXiv  Detail & Related papers  (2026-07-31T16:11:46Z)
• MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling [80.48332380100915]
  MiniCPM-SALA is a hybrid model that integrates the high-fidelity long-context modeling of sparse attention with the global efficiency of linear attention.<n>On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens.
  arXiv  Detail & Related papers  (2026-02-12T09:37:05Z)
• LINA: Linear Autoregressive Image Generative Models with Continuous Tokens [56.80443965097921]
  Autoregressive models with continuous tokens form a promising paradigm for visual generation, especially for text-to-image (T2I) synthesis.<n>We study how to design compute-efficient linear attention within this framework.<n>We present LINA, a simple and compute-efficient T2I model built entirely on linear attention, capable of generating high-fidelity 1024x1024 images from user instructions.
  arXiv  Detail & Related papers  (2026-01-30T06:44:33Z)
• Breaking the Bottleneck with DiffuApriel: High-Throughput Diffusion LMs with Mamba Backbone [6.76700377196741]
  We introduce DiffuApriel, a masked diffusion language model built on a bidirectional Mamba backbone.<n>Our results demonstrate that bidirectional state-space architectures serve as strong denoisers in masked diffusion LMs.
  arXiv  Detail & Related papers  (2025-11-19T23:23:49Z)
• Diffusion Language Models are Super Data Learners [61.721441061210896]
  When unique data is limited, diffusion language models (DLMs) consistently surpass autoregressive (AR) models by training for more epochs.<n>We attribute the gains to three compounding factors: (1) any-order modeling, (2) super-dense compute from iterative bidirectional denoising, and (3) built-in Monte Carlo augmentation.
  arXiv  Detail & Related papers  (2025-11-05T08:17:42Z)
• OmniSAT: Compact Action Token, Faster Auto Regression [70.70037017501357]
  We introduce an Omni Swift Action Tokenizer, which learns a compact, transferable action representation.<n>The resulting discrete tokenization shortens the training sequence by 6.8$times$, and lowers the target entropy.
  arXiv  Detail & Related papers  (2025-10-08T03:55:24Z)
• Spectral Compression Transformer with Line Pose Graph for Monocular 3D Human Pose Estimation [1.8999296421549172]
  We introduce the Spectral Compression Transformer (SCT) to reduce sequence length and accelerate computation.<n>The LPG generates skeletal position information that complements the input 2D joint positions.<n>Our model achieves state-of-the-art performance with improved computational efficiency.
  arXiv  Detail & Related papers  (2025-05-27T15:08:03Z)
• RainFusion: Adaptive Video Generation Acceleration via Multi-Dimensional Visual Redundancy [10.53687668536011]
  RainFusion exploits inherent sparsity nature in visual data to accelerate attention computation while preserving video quality.<n>Our proposed bf RainFusion is a plug-and-play method that can be seamlessly integrated into state-of-the-art 3D-attention video generation models.
  arXiv  Detail & Related papers  (2025-05-27T11:15:02Z)
• Systems and Algorithms for Convolutional Multi-Hybrid Language Models at Scale [68.6602625868888]
  We introduce convolutional multi-hybrid architectures, with a design grounded on two simple observations.<n>Operators in hybrid models can be tailored to token manipulation tasks such as in-context recall, multi-token recall, and compression.<n>We train end-to-end 1.2 to 2.9 times faster than optimized Transformers, and 1.1 to 1.4 times faster than previous generation hybrids.
  arXiv  Detail & Related papers  (2025-02-25T19:47:20Z)
• Large Scale Time-Series Representation Learning via Simultaneous Low and High Frequency Feature Bootstrapping [7.0064929761691745]
  We propose a non-contrastive self-supervised learning approach efficiently captures low and high-frequency time-varying features. Our method takes raw time series data as input and creates two different augmented views for two branches of the model. To demonstrate the robustness of our model we performed extensive experiments and ablation studies on five real-world time-series datasets.
  arXiv  Detail & Related papers  (2022-04-24T14:39:47Z)
• Chimera: Efficiently Training Large-Scale Neural Networks with Bidirectional Pipelines [23.62360119212958]
  This paper proposes Chimera, a novel pipeline parallelism scheme which combines bidirectional pipelines for efficiently training large-scale models.<n>Compared with the latest synchronous pipeline approach, Chimera reduces the number of bubbles by up to 50%.<n>For a GPT-2 model with 1.3 billion parameters running on 2,048 GPU nodes of the Piz Daint supercomputer, Chimera improves the training throughput by 1.16x-2.34x over the state-of-the-art synchronous and asynchronous pipeline approaches.
  arXiv  Detail & Related papers  (2021-07-14T18:16:20Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.