A Single-Line RoPE Tweak for Longer AI Videos

What it does RIFLEx, UltraViCo, and UltraImage are a family of plug-and-play methods that stretch the temporal or spatial limits of pre-trained diffusion transformers. The core trick is a surgical edit to the 1D Rotary Position Embedding (RoPE): the repository shows how to identify an intrinsic frequency in a pre-trained model and clamp it so the sinusoid stays within one period during extrapolation. For HunyuanVideo, this turns 5-second clips into 11-second ones; for CogVideoX-5B, 6 seconds become 12. The codebase spans multiple branches covering video (HunyuanVideo, CogVideoX, Wan2.1) and image (Flux, Qwen-Image) models.

The interesting bit The authors bill RIFLEx as a “free lunch,” and the README literally highlights the change as a single-line modification inside the standard RoPE function. That line—pinning one frequency component to 0.9 * 2 * torch.pi / L_test—is what keeps the attention mechanism from drifting out of distribution when the frame count grows. It is a rare case where a position-embedding band-aid generalizes to production-grade models without architectural surgery.

Key highlights

• Supports training-free inference and optional fine-tuned checkpoints (HunyuanVideo-RIFLEx, CogVideoX-RIFLEx).
• Branches cover Wan2.1, HunyuanVideo, Flux, and Qwen-Image, so the same idea applies to both video and high-resolution image generation.
• Multi-GPU inference branch exists for reproducing the project-page demos at full fidelity.
• The modification is model-agnostic to any diffusion transformer using RoPE; the repository provides a utility to find the intrinsic frequency index k.

Caveats

• The README notes that a few videos may show repetition in tail frames, so the conservative 0.9 multiplier deliberately leaves some headroom below a full period.
• Single-GPU Diffusers inference uses BitsAndBytesConfig quantization, which the authors warn can affect output quality compared with the multi-GPU reference.

Verdict Anyone running RoPE-based video or image diffusion models who needs longer or higher-resolution output without training from scratch should look here. If your model does not use RoPE, this repository is irrelevant.