Core idea. Tsinghua University and GigaAI jointly released the RhymeFlow framework, proposing an asynchronous denoising flow scheduling mechanism that significantly accelerates DiT (Diffusion Transformer)-based video generation models without retraining. The paper went online on arXiv on June 4 (2606.06309), with code open-sourced under Apache 2.0.
Technical breakthrough. Existing training-free acceleration methods (such as SVG, SAP, DiCache) mostly focus on "attention sparsification within a single denoising step," but still require every frame in the video to complete full dense denoising across all timesteps. RhymeFlow breaks this rigid constraint, dividing video frames into "key frames" and "non-key frames": key frames anchor semantic transitions, preserving dense stepwise denoising to keep structure intact; non-key frames skip predictable denoising steps progressively according to a "rhythmic sense," maintaining temporal consistency only through lightweight "latent-space trajectory projection" within 3D attention.
Performance data. On Wan 2.1, RhymeFlow surpasses SAP (24.45/0.730) with PSNR 26.29 and SSIM 0.783, achieving a 1.53× speedup; combined with SAP, the speed reaches 1.66×. On HunyuanVideo, used alone it achieves a 2.26× speedup, and stacked with SAP it reaches a staggering 2.60× speedup, with visual quality (PSNR/SSIM/LPIPS) comprehensively superior to baselines like SVG, EasyCache, DiCache, and VGDFR.
Perspective. RhymeFlow embodies an "orthogonal acceleration dimension" — not changing model weights, but reorganizing the inference timeline. This is highly complementary to sparse attention, KV-Cache, and speculative decoding. For DiT video models, the "compute behemoths," the "scheduling-is-optimization" line of thinking may be one of the most pragmatic paths for inference cost to keep falling in the second half of 2026.
Source: arXiv: 2606.06309 (2026-06-04); GitHub: Simon-Dcs/RhymeFlow