Stable Diffusion · LoRA
AceStep Refine Redmond
DPO-refined LoRA adapter for ACE-Step 1.5 Turbo that improves musicality, arrangement coherence, and vocal character in text-to-audio generation.
About this model
AceStep Refine Redmond is a DPO-refined LoRA adapter built on top of ACE-Step 1.5 Turbo (acestep-v15-turbo). The model was trained through a two-stage process: 75 epochs of large-dataset LoRA fine-tuning followed by DPO refinement on the resulting adapter, achieving approximately 70% win rate in blind A/B testing against the base model. The adapter uses LoRA rank 96 with alpha 192 and a learning rate of 8e-5.
The release includes two formats: a standard PEFT adapter for regular ACE-Step workflows, and a single-file ComfyUI-compatible LoRA export. For prompting and composition, the recommended language model is acestep-5Hz-lm-4B.
Known limitations include variable behavior with sparse prompts (less stable vocal timbre), potential texture noise or high-frequency harshness in very dense arrangements, and genre-specific generalization constraints due to the preference dataset used for DPO. The model is released under MIT license.
Learn more
Project signals
- 8 Hugging Face likes
Topics
peft · ace-step · lora · dpo · music-generation · audio-generation · text-to-audio · text2audio · acestep-v15-turbo · acestep-5hz-lm-4b