DaRealDahoodie/dark-blues-v2-xl-instrumental

Dark Blues v2 XL — Instrumental (Style/Quality)

LoRA adapter for ACE-Step 1.5 XL SFT focused on a narrow, high-quality dark blues / noir jazz cinematic instrumental style.

Companion adapter: DaRealDahoodie/dark-blues-v2-xl-vocals (vocal version)

Model Description

This adapter captures a moody, cinematic, jazz-inspired instrumental sound — think late-night reflection, controlled tension, high energy, balanced tone, and dark blues / noir jazz textures.

It was trained specifically on instrumental stems using the higher-rank "lora-quality" recipe for stronger stylistic adherence.

Primary use: Generating dark, atmospheric, late-night instrumental tracks with a noir/jazz/cinematic flavor.

Base Model

• Base: ACE-Step/acestep-v15-xl-sft (ACE-Step 1.5 XL SFT; also listed under ACE-Step/Ace-Step1.5)
• This is a PEFT LoRA adapter (rank 96). Load it on top of the official XL SFT checkpoint.

Training Details

• Dataset: 320 focused ~45-second clips (instrumental stems only) from the Dark_Blues_v2 collection.
• Recipe: lora-quality
• Rank (r): 96
• Alpha: 192
• Dropout: 0.08
• Target modules: q_proj, k_proj, v_proj, o_proj
• Epochs: Trained for 150 epochs; best checkpoint at epoch 110 (loss ~0.5913)
• Learning rate: 7.5e-5
• CFG dropout ratio (training): 0.15
• Other: batch size 1, grad accumulation 4, seed 42, precision bf16
• Training software: PiMPStudios/anvil-audio-v2

Note: This is a very stylistically narrow adapter (the source material was intentionally focused). It excels at this specific dark blues / noir jazz cinematic instrumental aesthetic rather than broad generalization.

How to Use

With the official ACE-Step Gradio / inference app

1. Load the base acestep-v15-xl-sft model.
2. Load this LoRA (point it at the folder or the adapter_model.safetensors).
3. Use prompts in the style of:
   cinematic, jazz, dark blues, noir jazz, moody instrumental, controlled tension, late night reflection

See the official ACE-Step 1.5 LoRA Training Tutorial for exact loading steps in the app.

Recommended prompting guidance (from training)

• Strong positive cues: cinematic, jazz, dark blues, noir jazz, moody instrumental, controlled tension, late night reflection, high energy, balanced tone
• Avoid in prompts: upbeat, happy, relaxing, bright, harsh treble, boomy low end, muddy mix

Limitations & Scope

• Optimized for instrumental (no vocals).
• Narrow stylistic focus — best results when your prompts lean into the late-night, moody, cinematic dark blues / noir jazz lane.
• Not intended as a general-purpose music LoRA.

Files

• adapter_model.safetensors — the LoRA weights
• adapter_config.json — PEFT configuration

License

MIT License

Copyright (c) 2026 PiMPStudios

See the LICENSE file for the full license text.

This is consistent with the base model ACE-Step/acestep-v15-xl-sft.

---

Created with: PiMPStudios/anvil-audio-v2 + ACE-Step 1.5 training stack (cloud job).
Best checkpoint selected: epoch_110_loss_0.5913 from the lora-quality run.