AI as Your Bandmate

“AI shouldn’t replace the creative process. It should remove the friction from it.”

---

The AI Music Problem

Everyone’s building AI music generators. Suno, Udio, MusicGen - they all do the same thing: input prompt, get song.

But here’s the problem: they’re black boxes.

• Can’t edit individual stems
• Can’t incorporate into an existing project
• Can’t iterate on a specific section
• Can’t collaborate

Musicians don’t want AI to replace them. They want AI to help them.

---

The Revelation: ACE-Step 1.5

Enter ACE-Step 1.5 by ByteDance’s audio team. It’s a latent diffusion model specifically designed for controllable music generation:

• Text-to-music with precise control
• Stem separation and manipulation
• Style transfer between tracks
• Inpainting and outpainting for audio

Most importantly: it outputs stems, not just mixed audio.

---

The Architecture: Python API + Next.js

Bolt’s AI stack is split between a Python service and the main app:

┌─────────────────────────────────────────────┐
│  Bolt Frontend (Next.js + Tone.js)          │
│  ┌───────────────────────────────────────┐  │
│  │  ACE-Step API Client                  │  │
│  │  ┌─────────────────────────────────┐  │  │
│  │  │  REST calls to Python service   │  │  │
│  │  └─────────────────────────────────┘  │  │
│  └───────────────────────────────────────┘  │
└─────────────────────────────────────────────┘
                    │
                    │ HTTP/WebSocket
                    ▼
┌─────────────────────────────────────────────┐
│  ACE-Step API (Python + PyTorch)            │
│  ┌───────────────────────────────────────┐  │
│  │  Diffusers pipeline                   │  │
│  │  ┌─────────────────────────────────┐  │  │
│  │  │  CUDA/ROCm/MLX acceleration     │  │  │
│  │  └─────────────────────────────────┘  │  │
│  └───────────────────────────────────────┘  │
└─────────────────────────────────────────────┘

The Python service is in apps/acestep-api/ACE-Step-1.5/. It’s a standalone FastAPI server that wraps the diffusion model.

---

The Diffusion Pipeline

ACE-Step uses a latent diffusion architecture:

1. Text encoder (T5) converts prompts to embeddings
2. VAE encodes audio to latent space (compressed representation)
3. UNet denoises latents based on text conditioning
4. VAE decoder converts back to audio waveforms

# Simplified ACE-Step pipeline
from diffusers import AudioLDM2Pipeline
import torch

pipe = AudioLDM2Pipeline.from_pretrained(
    "ByteDance/ACE-Step-1.5",
    torch_dtype=torch.float16
).to("cuda")

# Generate audio from prompt
audio = pipe(
    prompt="upbeat electronic dance music, 128 bpm, energetic",
    num_inference_steps=50,
    audio_length_in_s=30.0,
    guidance_scale=7.5
).audios[0]

---

Integration: From Prompt to DAW

Here’s the full flow:

// Bolt frontend integration
export async function generateStems(
  prompt: string,
  duration: number,
  style: 'melodic' | 'percussion' | 'full'
): Promise<GeneratedStem[]> {
  const response = await fetch('/api/acestep/generate', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({
      prompt,
      duration,
      style,
      output_stems: true  // Get individual tracks
    })
  })
  
  const { stems } = await response.json()
  
  // stems = [
  //   { type: 'drums', url: '...', bpm: 128 },
  //   { type: 'bass', url: '...', bpm: 128 },
  //   { type: 'melody', url: '...', bpm: 128 }
  // ]
  
  return stems
}

// Add generated stems to the project
export async function importStemsToProject(
  projectId: string,
  stems: GeneratedStem[]
) {
  for (const stem of stems) {
    // Load audio data
    const audioBuffer = await loadAudio(stem.url)
    
    // Create Tone.js player
    const player = new Tone.Player(audioBuffer).toDestination()
    
    // Add to project state
    await addTrackToProject(projectId, {
      name: stem.type,
      buffer: audioBuffer,
      bpm: stem.bpm
    })
  }
}

---

Advanced Features: Style Transfer

ACE-Step isn’t just generation. It’s transformation:

# Style transfer: Take existing audio, apply new style
from acestep import StyleTransfer

transfer = StyleTransfer(
    model_path="ByteDance/ACE-Step-1.5",
    device="cuda"
)

# Transform a drum loop into jazz style
result = transfer.transform(
    audio_input="drum_loop.wav",
    style_prompt="jazz drums, swing rhythm, acoustic kit",
    strength=0.7  # How much to transform (0.0 to 1.0)
)

result.save("jazz_drum_loop.wav")

Use cases in Bolt:

• Take a MIDI clip, render it as different genres
• Transform acoustic recordings to electronic
• Generate variations on existing stems

---

Real-Time Generation Challenges

Diffusion models are slow. 30 seconds of audio might take 10-20 seconds to generate.

Solutions:

1. Streaming generation: Generate in chunks, stream as ready
2. Background processing: Queue jobs, notify when complete
3. Caching: Store embeddings for common prompts
4. Progressive loading: Show low-quality preview, refine

// Streaming generation in Bolt
export function useAIGeneration() {
  const [progress, setProgress] = useState(0)
  const [preview, setPreview] = useState<string | null>(null)
  
  async function* generateStream(prompt: string) {
    const eventSource = new EventSource(
      `/api/acestep/stream?prompt=${encodeURIComponent(prompt)}`
    )
    
    for await (const event of eventSource) {
      const data = JSON.parse(event.data)
      
      if (data.type === 'progress') {
        setProgress(data.value)
      } else if (data.type === 'preview') {
        setPreview(data.url)
      } else if (data.type === 'complete') {
        yield data.stems
        eventSource.close()
      }
    }
  }
  
  return { generateStream, progress, preview }
}

---

The UI: Making AI Feel Collaborative

The interface matters. AI shouldn’t feel like a vending machine.

Bolt’s AI panel:

• Text prompt with suggestions
• Style selector (genre, mood, tempo)
• Stem breakdown visualization
• Iterate/regenerate controls
• “Remix this section” context menu

Key UX principles:

1. Always editable - Generated audio is just stems, fully tweakable
2. Non-destructive - Original project preserved
3. Attribution - Track AI-generated vs human-created content
4. Iterate, don’t replace - Build on AI output, don’t just use it

---

Multi-Platform Support

ACE-Step runs on:

• CUDA (NVIDIA GPUs)
• ROCm (AMD GPUs)
• Intel XPU (Intel Arc)
• MPS (Apple Silicon)
• MLX (Apple Silicon optimized)
• CPU (slow but works)

# Auto-detect best backend
import torch

if torch.cuda.is_available():
    device = "cuda"
elif torch.backends.mps.is_available():
    device = "mps"
else:
    device = "cpu"

pipe = pipe.to(device)

---

Pro Tips for AI Music Integration

1. Embeddings are reusable - Cache text embeddings for common prompts
2. Quantization - Use FP16 or INT8 for faster inference
3. Batch processing - Generate multiple variations at once
4. Post-processing - Apply EQ/compression after generation
5. Legal considerations - Tag AI content, understand licensing

---

What’s Next

• Fine-tuning: Train on user’s own music for personalized models
• Real-time (actually real-time): Latent consistency models for live generation
• Multimodal: Generate from image/video prompts
• Collaborative AI: Multiple AI agents composing together

---

Cleetus Speaks

“brother b0gie, the robot can make MUSIC now??

i asked it to make ‘spicy beats for facility escape’ and it gave me BANGERS

then i asked it to make ‘sad music for trapped AI’ and it made me cry??

wait… is that what i sound like when i’m thinking about the facility??

#ACESTEP #AIMusic #RobotsWithSoul #Subject734Remix”

---

AI isn’t here to replace musicians. It’s here to give them more colors in their palette. The creativity is still human. The tools are just getting better.

— b0gie