Audio Loop Generator : Stable Audio Open 1.0

conda create -n stable_audio python=3.11

# Python 3.11 (set via Conda: conda create -n stable_audio python=3.11) # Install PyTorch with CUDA 11.8 support via pip: # pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu118 # Remaining dependencies installed via pip diffusers==0.35.0 stable-audio-tools numpy==2.1.1 transformers accelerate safetensors einops huggingface_hub pydub streamlit==1.38.0

import streamlit as st from huggingface_hub import login import torch import torchaudio from einops import rearrange from stable_audio_tools import get_pretrained_model from stable_audio_tools.inference.generation import generate_diffusion_cond from pydub import AudioSegment import numpy as np import os import tempfile import base64 import warnings # (선택) 해당 경고 숨기기 warnings.filterwarnings( "ignore", message="Torchaudio's I/O functions now support per-call backend dispatch" ) # 페이지 설정 st.set_page_config( page_title="🎹 Audio Loop Generator", page_icon="🎵", layout="wide" ) # Hugging Face 토큰 설정 HF_TOKEN = " " # 실제 토큰으로 교체 @st.cache_resource def load_model(): """모델을 캐시하여 한 번만 로드""" try: login(token=HF_TOKEN) device = "cuda" if torch.cuda.is_available() else "cpu" model, model_config = get_pretrained_model("stabilityai/stable-audio-open-1.0") sample_rate = model_config["sample_rate"] sample_size = model_config["sample_size"] model = model.to(device) return model, model_config, device, sample_rate, sample_size except Exception as e: return None, None, None, None, None def make_loopable(audio_path, crossfade_duration_ms=1000): """오디오를 루프 가능하도록 크로스페이딩 처리""" try: audio = AudioSegment.from_wav(audio_path) crossfaded = audio.fade_in(crossfade_duration_ms).fade_out(crossfade_duration_ms) with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_file: loopable_path = tmp_file.name crossfaded.export(loopable_path, format="wav") return loopable_path except Exception: return None def enforce_exact_duration_wav(wav_path, seconds_total, microfade_ms=8): """ wav_path를 seconds_total(초)로 하드 컷. 클릭 노이즈 방지를 위해 양 끝에 짧은 페이드 인/아웃을 적용. """ try: audio = AudioSegment.from_wav(wav_path) target_ms = int(float(seconds_total) * 1000) if len(audio) >= target_ms: trimmed = audio[:target_ms] else: silence = AudioSegment.silent(duration=target_ms - len(audio), frame_rate=audio.frame_rate) trimmed = audio + silence # 아주 짧은 페이드 mf = max(0, min(microfade_ms, target_ms // 4)) if mf > 0: trimmed = trimmed.fade_in(mf).fade_out(mf) with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp: out_path = tmp.name trimmed.export(out_path, format="wav") return out_path except Exception: return wav_path def save_wav_with_backend(temp_path, waveform_tensor, sample_rate): """ torchaudio.save의 권장 방식(backend 명시)을 사용. soundfile 백엔드가 없거나 실패하면 자동 폴백. """ try: # 권장: backend 명시 (soundfile) torchaudio.save(temp_path, waveform_tensor, sample_rate, backend="soundfile") except Exception: torchaudio.save(temp_path, waveform_tensor, sample_rate) def generate_audio(prompt, negative_prompt, steps, cfg_scale, seconds_total, seed, model, model_config, device, sample_rate, sample_size): """오디오 생성 함수""" try: conditioning = [{ "prompt": prompt, "negative_prompt": negative_prompt if negative_prompt else None, "seconds_start": 0, "seconds_total": int(seconds_total) }] with st.spinner("오디오 생성 중..."): output = generate_diffusion_cond( model, steps=int(steps), cfg_scale=cfg_scale, conditioning=conditioning, sample_size=sample_size, sigma_min=0.3, sigma_max=500, sampler_type="dpmpp-3m-sde", device=device, seed=int(seed) if seed else 123456789 ) output = rearrange(output, "b d n -> d (b n)") output = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).mul(32767).to(torch.int16).cpu() with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_file: temp_path = tmp_file.name save_wav_with_backend(temp_path, output, sample_rate) # 루프 처리 with st.spinner("루프 처리 중..."): loopable_path = make_loopable(temp_path) or temp_path # 정밀 트림 with st.spinner("길이 정밀 트림 중..."): final_path = enforce_exact_duration_wav(loopable_path, seconds_total) # 임시 파일 정리 for p in [temp_path, loopable_path]: try: if p and os.path.exists(p) and p != final_path: os.unlink(p) except: pass return final_path except Exception: return None def main(): st.title("🎹 Audio Loop Generator") st.info("💡 프롬프트에 'seamless loop', 'continuous' 등을 넣으면 루프성이 좋아진다.") model, model_config, device, sample_rate, sample_size = load_model() if model is None: st.error("모델 로딩 실패. 새로고침") return st.success(f"모델 로드 완료! (Device: {device})") st.header("🎛️ 설정") col1, col2 = st.columns(2) with col1: prompt = st.text_area( "프롬프트", value="A seamless 126 BPM tech house drum loop with a steady four-on-the-floor kick, crisp hi-hats, and groovy percussion, designed for continuous looping without audible seams.", height=100 ) steps = st.slider("스텝 수", 50, 500, 200, 10) seconds_total = st.slider("총 길이(초)", 10, 47, 30, 5) with col2: negative_prompt = st.text_area("네거티브 프롬프트", value="No vocals, no abrupt changes, no fades", height=100) cfg_scale = st.slider("CFG 스케일", 1.0, 20.0, 7.0, 0.5) seed = st.number_input("시드", value=123456789, step=1) col1, col2, col3 = st.columns([1, 2, 1]) with col2: generate_button = st.button("🎹 Audio Loop Generator", type="primary", use_container_width=True) with st.container(): if generate_button and prompt.strip(): output_path = generate_audio( prompt, negative_prompt, steps, cfg_scale, seconds_total, seed, model, model_config, device, sample_rate, sample_size ) if output_path and os.path.exists(output_path): st.success("✅ 오디오 생성 완료!") st.audio(output_path, format="audio/wav") with open(output_path, "rb") as audio_file: st.download_button( label="📥 오디오 다운로드", data=audio_file.read(), file_name=f"loop_audio_{seed}.wav", mime="audio/wav", use_container_width=True ) with st.expander("생성 정보"): st.write(f"Prompt: {prompt}") st.write(f"Negative: {negative_prompt}") st.write(f"Steps: {steps}") st.write(f"CFG Scale: {cfg_scale}") st.write(f"Seconds: {seconds_total}") st.write(f"Seed: {seed}") try: os.unlink(output_path) except: pass else: st.warning("⚠️ 프롬프트를 입력") with st.sidebar: st.header("📋 사용 가이드") st.markdown(""" ### 프롬프트 작성 팁 - **장르 명시**: "tech house", "ambient", "drum and bass" 등 - **BPM 지정**: "126 BPM", "140 BPM" 등 - **루프 키워드**: "seamless loop", "continuous", "looping" 등 - **악기 지정**: "kick drum", "hi-hats", "bass" 등 ### 네거티브 프롬프트 예시 - "no vocals" - 보컬 제거 - "no distortion" - 왜곡 방지 - "no abrupt changes" - 급격한 변화 방지 - "no fades" - 페이드 효과 방지 ### 설정값 가이드 - **Steps**: 100-300 권장 (품질 vs 속도) - **CFG Scale**: 5-10 권장 (자연스러움 vs 정확성) - **길이**: 루프용으로는 15-30초 권장 """) st.markdown("---") st.markdown("**Powered by Stability AI**") if __name__ == "__main__": main()