Modelpublic

DisturbingTheField/ACE-Step-v1.5-raspy-vocal-and-instrumental-5-LoRAs

sourceHugging Facemitupdated 7mo agoView on Hugging Face

3likes70downloads

Model Card

Date: 2/20/2026

Rename these LoRA files to only adaptermodel.safetensors when using. The same adapterconfig.json can be used for all LoRA files.

LoRA file: malevocalsadapter_model.safetensors

Sounds: Me singing when I had a cold and a very hoarse voice Training: 7 self recorded wav files, 58 MB, 1200 epochs

LoRA file: instrumentaladaptermodel.safetensors

Sounds: Instrumental songs made by myself. Include electric guitar, distorted guitar, bass, drums, piano, synth etc. Training: 21 files, 766 MB, 800 epochs

Both trained with acestep-v15-base and acestep-5Hz-lm-4B. Dataset, preprocessed tensors with both, and training with only acestep-v15-base. Both trained on a laptop, Nvidia RTX 3080 Ti with 16 GB VRAM. When making songs I believe I used acestep-v15-turbo with acestep-5Hz-lm-1.7B.

Merged LoRA files

Use the Python script MERGE-LORA.py.txt, read some more info in the script. I made these 3: Strength: voc 0.8 inst 0.8 (adaptermodel.safetensors) Strength: voc 0.6 inst 1.4 (voc06inst14__adaptermodel.safetensors) Strength: voc 1.4 inst 0.6 (voc14inst06_adaptermodel.safetensors) You can use the same adapter_config.json with all LoRAs.

Description

There are 2 LoRA adapters here. The first is pure vocal, trained on my own voice when I had a cold and a very hoarse voice with uncontrollable pitch. The second is pure instrumental, trained on 21 instrumental tracks that I made myself. The music styles vary: rock, acoustic guitar, distorted guitar, ambient, etc. Instruments include clean guitar, distorted guitar, bass, drums, piano, synth, and more.

Use these two as a kind of filter — they work best with low LoRA Scale values between 0.2 and 0.7. Also test them with “Think” both on and off, and try varying the LM Temperature and LM CFG Scale values.

There is also a Python script included called MERGE-LORA.py.txt, which can be used to combine two LoRA adapters. You set the strength for each of them. The two LoRA adapters you combine must contain the same layers — typically meaning they were created from the same base model. The script checks this before generating the merged version. See the script for more information if you want to use it.

Additionally, there are 3 more LoRA adapters included, which are simply three different combinations of the two main LoRA adapters. These provide both vocal and instrumental effects.

I’ve included some demo MP3 files as well. These are not necessarily polished songs, but rather examples so you can hear the kinds of sounds/effects you can achieve. There are a lot of possibilities here, so I can’t test everything — you’ll just have to experiment yourself.

Again, think of these two main LoRA adapters more as filters that adjust aspects of the vocals and instruments. The vocal LoRA adapter will affect both the vocals, the instruments, and the overall song. Songs become calmer, almost sadder, when used with stronger values.

You can store everything in one folder, but make sure you have a JSON file named “adapterconfig.json” and a LoRA file named “adaptermodel.safetensors”. They must have these names in order to be loaded properly. So rename the LoRA files according to which one you want to load.

3 captions for testing:

—90s dance feel-good vibe: Upbeat, feel-good dance music inspired by the smooth European club sound of the late 1990s. Male singer. The groove is driven by steady four-on-the-floor drums and a warm, rounded bassline that locks tightly with the rhythm. Clean electric guitar adds light, funky chord stabs and rhythmic accents, giving the track a fresh, organic touch against a backdrop of shimmering synth pads, bright keyboard hooks, and subtle electronic textures. The production is polished and uplifting, blending disco-influenced grooves with pop sensibility and dancefloor energy. The overall vibe is sunny, nostalgic, and effortlessly catchy—music designed to feel carefree, stylish, and movement-driven. Male singer. Funky clean guitar chords. Synth pads soft in the background.

—Rock: Melodic British-style rock with a polished yet organic sound, driven by clean, articulate lead guitar lines and a steady, radio-friendly mid-tempo groove. The arrangement features tight rhythm guitar, supportive bass lines, crisp drums, and subtle dynamic builds. The vocal delivery comes from a male singer with a dark, slightly raspy voice, combining understated intensity with a conversational, storytelling approach. He has a deep, dark baritone with a gravelly, rough-edged texture. There’s a raw, raspy quality to his voice. The overall feel is catchy and accessible, blending classic rock sensibility with pop structure and memorable hooks. High-fidelity, studio-polished.

—Electronica, ambient, dance etc: Music combines dark, moody atmospheres with melodic electronic pop. Male voice. He has a deep, dark baritone with a gravelly, rough-edged texture. There’s a raw, raspy quality to his voice. Synthesizers dominate the sound, layering rich textures, pulsating basslines, and hypnotic arpeggios. The vocals are expressive, sometimes melancholic or brooding, and often carry a sense of intimacy or vulnerability. Guitar parts occasionally cut through, adding grit or accentuating climactic moments. The song explore themes of love, desire, pain, and introspection, often tinged with darkness. Effects like reverb, delay, and subtle distortion enhance the emotional atmosphere, giving the music a cinematic quality. Syth adds melody hooks. The overall sound is both danceable and haunting, blending electronic sophistication with raw human emotion. The lyrics are almost spoken sometimes, with deep dark voice.

Lyrics for testing:

[Intro]

[Verse 1] Sun is rising slowly, light upon the floor Coffee’s on the table, no alarms, no chores Kids are laughing softly, in the morning glow Nothing on the schedule, nowhere we need to go

[Chorus] Oh, Sundays feel like heaven, hearts are running free Time to love, time to linger, just my family and me The world can wait a little, we’ve got our own parade Oh, Sundays feel like magic, every moment we have made

[Verse 2] Stories in the kitchen, songs drift through the air Moments like these linger, precious and rare The clock is just a number, the hours drift away Wrapped up in each other, there’s nothing left to say

[Outro]

Model Card for Model ID:

5 LoRA adapters with raspy male vocals, instrumental and mix of both for ACE-Step/Ace-Step1.5

Model Description: If yu use "startgradioui.bat" then edit the file: set INITSERVICE=--initservice false So you get access to the LoRA loading part in the Web UI.

Model Sources: https://huggingface.co/ACE-Step/Ace-Step1.5

Uses: 5 LoRA adapters with raspy male vocals, instrumental and mix of both for ACE-Step/Ace-Step1.5

Training Details:

LoRA file: malevocalsadapter_model.safetensors Sounds: My singing when I had a cold and a very hoarse voice Training: 7 self recorded wav files, 58 MB, 1200 epochs

LoRA file: instrumentaladaptermodel.safetensors Sounds: Instrumental songs made by myself. Include electric guitar, distorted guitar, bass, drums, piano, synth etc. Training: 21 files, 766 MB, 800 epochs