ntc-ai/yue2-particle-sliders

YuE2 particle sliders

Generate a same-caption, same-lyrics, same-seed Off/On comparison with any of 16 voice and genre controls. Strength 0 bypasses the adapter; +1 is the trained positive endpoint. Browse all 64 release comparisons without consuming GPU quota in the listening tab.

Model repo, weights and samples · Native loading · MiniMax Music 3 demo

These are the final EMA checkpoints from the 1,200-update routed-particle campaign. The native nonlinear adapters run only during AR planning/composition. The acoustic stage and VAE use frozen base weights. They require the included native loader and cannot be applied using a generic LoRA node.

Song length defaults to Automatic: the native YuE2 sampler decides when to emit its ending token. No seconds-based cap is applied in this mode. Choose Limited to set an optional maximum duration. Off and On use the same caption, lyrics and seed, but can end at different times.

YuE2 runs in BF16 through its native PyTorch CUDA-graph decoder. Graph capture includes the native particle projections; the adapter is removed before acoustic synthesis. Both the base model and VAE are loaded at startup for ZeroGPU. Requests are serialized; output paths are unique. Automatic mode retains the upstream 9,000-token safety guard (about six minutes). The GPU reservation is estimated separately and capped at 120 seconds after the Spaces hardware duration multiplier; it never changes the sampling limit. Long requests remain subject to ZeroGPU's execution-time and quota limits. Generation details report the active limit, semantic timing and any truncation. A live automatic-generation check produced a naturally ended 58.20-second song in 28.23 seconds with a 34-second scheduler request. Audio quality and lyric preservation are not guaranteed by the training diagnostics.

yue2/ is the unmodified upstream inference package from revision ef1936f2ee39fe8de486a0f47a481c95f8d4da87, bundled with its licenses. Bundling avoids installing the upstream wheel's exact Hub pin into the platform-managed environment. Gradio is pinned to 6.17.3, the newest release compatible with the Hub range required by YuE2's tested Transformers 4.57.6; Gradio 6.18 and later require Hub 1.x. slider_runtime.py contains the exact native inference classes extracted from training source; source-provenance.json records their source hashes.

Base checkpoints: YuE2-3B and YuE2-Vae. Model terms are in upstream-licenses/MODEL_LICENSE; upstream inference code uses Apache-2.0. Adapter checkpoints use CC BY-NC 4.0.