File size: 2,176 Bytes
43fb1df
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1472005
 
 
 
43fb1df
 
 
 
 
 
1472005
 
43fb1df
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
---
language:
- en
license: mit
pipeline_tag: text-to-audio
tags:
- ACE-Step
- LoRA
- DPO
- music-generation
- audio-generation
- text-to-audio
- text2audio
- PEFT
- acestep-v15-turbo
- acestep-5Hz-lm-4B
base_model:
- ACE-Step/Ace-Step1.5
library_name: peft
widget:
- text: "Showcase reel"
  output:
    url: showcase-training-chapter-v3.mp4
---

# AceStep_Refine_Redmond

I'm grateful for the GPU time from Redmond.AI that allowed me to make this model!

<Gallery />

## Overview
AceStep_Refine_Redmond is a DPO-refined LoRA adapter for ACE-Step 1.5 Turbo, focused on improving musicality, arrangement coherence, and vocal character in practical generation workflows.

This release includes:
- `standard/` (PEFT adapter for regular ACE-Step loading)
- `comfyui/` (single-file ComfyUI-compatible LoRA export)

## Compatibility
- DiT used: `acestep-v15-turbo`
- Recommended LM for prompting/composition: `acestep-5Hz-lm-4B`
- `standard/` works in regular ACE-Step workflows.
- `comfyui/` is the converted single-file LoRA for ComfyUI.

## What Changed vs Base
In blind A/B testing against the base reference, this refinement achieved about **70% win rate**.
The blind test votes were collected from different users.

Training summary (final DPO refinement stage):
- Base checkpoint: `acestep-v15-turbo`
- Adapter type: LoRA
- Rank / Alpha: `96 / 192`
- Learning rate: `8e-5`
- Training path: large-dataset LoRA fine-tune for `75` epochs, then DPO refinement on top of that adapter
- Epoch config: up to `81` in the DPO stage (resumed from the previous epoch-75 adapter)

## Known Limitations
- Behavior can still vary by prompt style; some sparse prompts may produce less stable vocal timbre.
- Very dense arrangements can introduce texture noise or high-frequency harshness in some generations.
- This adapter is tuned on a specific preference dataset and may not generalize equally across all genres.

## Responsible Use
- Do not use this model to imitate or impersonate real artists without permission.
- Respect copyright, voice rights, and local regulations when generating and publishing audio.
- Review outputs before public release, especially in commercial workflows.