Tempolor v3.5 is based on a diffusion model and continuous music-audio representations, supporting the generation of 44.1kHz stereo high-quality music. In its technical roadmap, this version evolves from a paradigm emphasizing semantic-skeleton modeling toward a high-fidelity paradigm centered on continuous acoustic representations, markedly improving sound texture, spatial layering, dynamics and the overall naturalness and completeness of the listening experience.

By introducing controllable-generation capabilities such as ControlNet and Inpaint / Repaint, Tempolor v3.5 further expands editability and controllability in scenarios like melody control, lyric editing and local repainting, providing a technical foundation for finer music creation and interactive editing.

In terms of inference efficiency, the real-time factor is below 0.1, with full-song generation speed at the industry-leading level.

Tempolor v3.5 has a clear advantage in fine listening detail and atmospheric expression. The model more naturally restores reverb tails, dynamic swells, textural layering and spatial depth, making results more immersive in emotion-driven content such as lyrical, ethereal, suspenseful and epic pieces.

Compared with the previous two generations, 3.5 cares not only about "writing it right" but about "sounding good," making it better suited to film scoring, atmospheric music and brand mood music — scenarios with higher demands on listening completeness.

Especially in vocals, it delivers excellent vocal quality across singing skill, vocal timbre, vocal fidelity and lyric clarity.

* Based on NVIDIA RTX 4090

Comparison of time to generate 120s of music audio (inference on an Nvidia L20 GPU)