AceStep Refine Redmond: DPO LoRA for Better Music Generation

TL;DR

• DPO-refined LoRA adapter for ACE-Step 1.5 Turbo with 70% win rate in blind A/B tests
• Two-stage training: 75 epochs LoRA fine-tune + DPO refinement (LoRA rank 96, alpha 192)
• Works in standard ACE-Step and ComfyUI workflows; MIT licensed

What This Model Does

AceStep Refine Redmond is a LoRA adapter that improves the output quality of ACE-Step 1.5 Turbo (acestep-v15-turbo) for text-to-audio generation. The refinement targets three specific areas: musicality, arrangement coherence, and vocal character. It was developed using GPU compute sponsored by Redmond.AI.

The adapter is compatible with the ACE-Step pipeline and comes in two formats:

• standard/ — PEFT adapter for regular ACE-Step loading
• comfyui/ — Single-file LoRA export for ComfyUI workflows

For best prompting results, the author recommends using the acestep-5Hz-lm-4B language model for composition and prompt engineering.

Training Details

The model was trained in two stages:

1. Large-dataset LoRA fine-tuning: 75 epochs on the base acestep-v15-turbo checkpoint
2. DPO refinement: Direct Preference Optimization on top of the epoch-75 adapter, continued up to epoch 81

LoRA configuration:

• Rank: 96
• Alpha: 192
• Learning rate: 8e-5
• Adapter type: LoRA (PEFT)

In blind A/B testing against the base ACE-Step 1.5 Turbo, this refinement achieved approximately 70% win rate across multiple users.

How to Use

Standard ACE-Step (PEFT)

from acestep.pipeline import ACEStepPipeline

pipe = ACEStepPipeline.from_pretrained(
    "ACE-Step/Ace-Step1.5",
    adapter_name_or_path="artificialguybr/AceStep_Refine_Redmond",
    subfolder="standard"
)

ComfyUI

Place the file from comfyui/ in your ComfyUI models/loras/ directory and load it as a standard LoRA.

Recommended Settings

• Base model: ACE-Step/Ace-Step1.5 (acestep-v15-turbo)
• Prompt LM: acestep-5Hz-lm-4B
• Use descriptive prompts for best vocal stability

Known Limitations

• Sparse prompts: May produce less stable vocal timbre with minimal prompt descriptions
• Dense arrangements: Very complex arrangements can introduce texture noise or high-frequency harshness
• Genre generalization: Tuned on a specific preference dataset; performance may vary across genres

Responsible Use

• Do not use to imitate or impersonate real artists without permission
• Respect copyright, voice rights, and local regulations
• Review outputs before public release, especially in commercial workflows

License

MIT License — see the Hugging Face model card for details.

Common Errors and Fixes

Error | Cause | Fix | Source

|-------|-------|-----|--------

LoRA not loading in ComfyUI | Wrong file path or format | Use the comfyui/ single-file export, place in models/loras/ | Model card Vocal instability | Sparse/underspecified prompt | Use descriptive prompts with acestep-5Hz-lm-4B for composition | Model card High-frequency harshness | Very dense arrangement prompts | Simplify arrangement density in prompt | Model card

FAQ

Q: What base model does this LoRA require? A: ACE-Step/Ace-Step1.5 (acestep-v15-turbo). It will not work with other ACE-Step versions.

Q: Can I merge this LoRA into the base model? A: Yes, standard PEFT merging tools can merge the standard/ adapter into the base model weights.

Q: Is this model commercial-use friendly? A: Yes, MIT license permits commercial use. Review the responsible use guidelines regarding artist impersonation and voice rights.

Q: What's the difference between the standard and comfyui folders? A: standard/ is a PEFT adapter directory for programmatic loading. comfyui/ is a single .safetensors file for drag-and-drop ComfyUI use.

Sources

• Hugging Face Model Card — Primary source for all training details, usage instructions, and limitations
• ArtificialGuyBR Portfolio — Author's model portfolio and other projects
• FindGoodPrompt — Curated prompts for image and language models
• JustGoodAI — Curated AI tools directory