ailm32442/SoulX-Singer-onnx-fp32

SoulX-Singer ONNX (fp32)

FP32 ONNX export of SoulX-Singer (zero-shot singing voice synthesis, Soul AI Lab), split into 9 ONNX graphs plus 7 external-weight .onnx.data files.

This is the high-fidelity tier: diff_step_dml.onnx and vocoder_dml.onnx are kept in fp32 (the fp16 tier is 1.44 GB, this one is 2.66 GiB). The encoder graphs are byte-identical to the fp16 export.

• Upstream weights: Soul-AILab/SoulX-Singer 鈥?Apache-2.0
• ONNX split export: SXSEditor (DirectML build)
• Bytes mirrored from: ModelScope syxppp/SoulX-Singer-onnx-directml 鈥?byte-identical (鍥藉唴鍙敤鍘熶粨鐩磋繛)
• Redistributed under the Apache License 2.0. See NOTICE in the consuming project for attribution.

Files

| note_text_encoder.onnx | 433 | 087db35205ce93dae9e2c65e0628511266ffbc21beb2c128d80793c150d66714 | | note_text_encoder.onnx.data | 6144000 | 998889d7daff9cd50774abf23282d1acad01d2c753db3040b156d1a53b178e79 | | note_pitch_encoder.onnx | 433 | 90f9ef1cd74e532e1cc89802888bedcf311bd835460b09e27c3eef1be9489bdb | | note_pitch_encoder.onnx.data | 524288 | 20ddcaaea04315e5a46eb11d8cacb6dd6ecf6a9688ebc3af019ab7b27ee7b9c6 | | note_type_encoder.onnx | 432 | 869d4d50a0ff5566346e72cf4fb36c110a52e10ef7292c53aecd008c3f74d594 | | note_type_encoder.onnx.data | 524288 | becf918cd5e23b25783ae42480c13bfa04bf38b84a4075d7c4273befcb3af4a0 | | f0_encoder.onnx | 425 | f63a5f8da66af2a50b6c3463f8e8a0b52022f6f14bbb79d7364149da2c576a0f | | f0_encoder.onnx.data | 739328 | 8c3f3c514b1bf78f723b8fa785ed2f7dd5cb994ca2c1349f22c54ea4909b0f4b | | preflow.onnx | 145261 | af0680d7356ba1131deb5a1801971934344f7ccd5394c875a612f8a66c8e1a60 | | preflow.onnx.data | 16973824 | ce2c424a3a22a22a1aff172767baacc3e6796df0eb83f979596d522d87d2422f | | cond_emb.onnx | 689 | 6825d68ea4d42bc83ca66dacf3a6440f5f2593af2a8ca40f956fdcde19559e54 | | cond_emb.onnx.data | 2162688 | fbfa95df1617d8dc12c1544c788be65d5bb08b365a68041c69953b02ac066dbe | | mel_transform.onnx | 11263 | 5652b9b8a46207413db3d2ff65139d78ab502d60fa7eb3a1a660ec8296d1d885 | | mel_transform.onnx.data | 499712 | 2d1e18edf37041270f999b15bbfd08c19cccca723a6111425ae9a8a228c87927 | | diff_step_dml.onnx | 1772509288 | bc6e95474793a0eb7dbf8ea4c7cd3820593f6bd5d136f222ea58684b64b07c13 | | vocoder_dml.onnx | 1053635812 | 8df53cf50ff7f91208fe613f5f777690a2b9ddb9d17fa59bc6efa0fdbf1e8535 |

Total: 2.66 GiB (2853872164 bytes)

Usage

The .onnx graphs locate their weights through relative paths recorded inside the graph, so all 16 files must stay in the same directory. Load with ONNX Runtime (DirectML / CUDA / CPU); diff_step_dml.onnx is the flow-matching step graph and is invoked repeatedly (16鈥?2 steps).