✨ Song Crafting Studio
🎛 Generation Controls & ACE-Step Blueprint
🧠 Thinking AI Reasoning
0.6B LLM chain-of-thought analysis
🎵 Instrumental Only
Generate pure instrumental mix
⚡ CPU Offload
Reduce VRAM pressure on T4
Standard: 1.0 - 2.5
📋 Order Blueprint Summary
Task: Text2Music
Turbo 8-Steps
180s @ 115 BPM
🧠 Thinking: ON
🇹🇷 Vocal: tr
⚡ Pure VRAM
Key: Auto
🎤 1. Reference Audio Source
Audio-to-Audio Conditioning
Choose a previously synthesized song from your library or upload an MP3/WAV reference to guide the melody and harmony.
🎛 2. Cover & Style Remix Blueprint
Lower = Free Remix • Higher = Strict Melody
🧠 Thinking AI
0.6B LLM chord analysis
🎵 Instrumental
Make cover beat only
📋 Cover Order Blueprint
Task: Cover / Audio2Audio
Task: Cover
Similarity: 60%
Turbo 8-Steps
🧠 Thinking: ON
180s @ 115 BPM
📥 Bulk .JSON & .ZIP Batch Ingestion
Drop your song .json file(s) or a .zip archive. The system automatically extracts musical prompt, lyrics, BPM, language, and Song ID.
Drag & Drop song .JSON or .ZIP files here
or click to browse your computer (select one or multiple files)
📋 Active Generation Queue
| ID | Title | Duration | BPM | Steps | Seed | Status | Actions | |
|---|---|---|---|---|---|---|---|---|
| No jobs in queue. | ||||||||
🎵 Completed Master Tracks & Packages
| Song ID | Title & Artist | Duration | BPM | Model | Completed At | Playback, Remix & Export | |
|---|---|---|---|---|---|---|---|
| No completed tracks yet. | |||||||
⚡ Memory & Hardware Optimization
Configure how the remote Google Colab Tesla T4 GPU interacts with host CPU RAM and VRAM.
💾 Model CPU Offloading (offload_to_cpu)
Offloads inactive model layers to system RAM to prevent CUDA Out-of-Memory.
🔄 DiT Model CPU Offloading (offload_dit_to_cpu)
Moves DiT diffusion network to CPU while the 0.6B LLM generates lyrics and chords.
⚡ Flash Attention Acceleration
Uses optimized flash attention kernel for faster diffusion passes.
🧠 Thinking Model & AI Reasoning Defaults
The ACE-Step 0.6B language model analyzes your caption and prompt to build musical blueprints before diffusion.
🧠 Enable Thinking AI by Default
Automatically enables Chain-of-Thought reasoning for all new standard and cover tracks.
☁ Google Colab Cloud Multi-Account
Seamlessly switch between Google accounts to bypass free-tier T4 GPU quota limits.
⚡ Automated Power Management
⏹ Auto-Power Off Colab on Queue Completion
Automatically terminates Colab session when all queue songs are finished to save compute units.
⚡ Auto-Wake Colab on Task Assignment
Automatically spins up Google Colab GPU runtime the moment a new track or cover is queued.
⚡ Vast.ai Dedicated GPU Cloud Profile
⚪ Disconnected (Standby)
Connect a dedicated rented GPU instance from Vast.ai (e.g. RTX 3090, 4090, A100) to generate music and FLUX.2 Klein images in pure high-speed VRAM without Google Colab disconnects.
🔒 Auto-detected: ~/.ssh/id_ed25519
Or paste private key text directly (optional)
🎼 Production & Musical Defaults
Turbo: 8-step diffusion in pure VRAM.
Base: 32-step high fidelity.
XL: 4-billion parameter model for intricate arrangements.
Base: 32-step high fidelity.
XL: 4-billion parameter model for intricate arrangements.
Language Fallback Behavior: Only used if an imported prompt or JSON lacks language tags. If you prompt Arabic music, ACE-Step synthesizes Arabic vocals without clashing with Turkish.
Target song length in seconds (180s = 3 minutes). Shorter tracks render faster with lower VRAM.
Set a numeric tempo (e.g. 115) or type auto to auto-extract tempo from the prompt or let AI choose.
📦 Metadata, ID3 Tags & Storage
Fallback ID3 tag. If your imported JSON batch contains an 'artist', that value overrides this default.
Fallback ID3 tag. If your imported JSON batch contains an 'album', that value overrides this default.
📜 Structured Diagnostic Event Console
[System] Connecting to Studio event stream...
👥 Invite User & Grant Timed Access
Grant someone private access to generate music. You decide how long their access lasts. Their generations, audio library, and downloads remain strictly isolated from yours.
🔔 Telegram Alerts Bot
ACTIVE
Your personal Telegram bot notifies you in real-time when batch jobs finish and when GPUs change state.
Target Chat ID:
Fetching...
Batch Complete Alerts:
Enabled (2+ Songs)
Vast.ai Connect/Disconnect:
Enabled (Debounced)
👥 User Accounts & Activity
| User | Role | Status | Duration / Expiry | Active Jobs | Total Songs | Last Active | Actions |
|---|---|---|---|---|---|---|---|
| Loading users... | |||||||
📁 Batch JSON Input
Drop JSON batch file here or click to browse
Supports structured ImageBatchJsonSpec format
⚙️ Generation Parameters
📐 Target Resolution Overrides
💾 Saved
×
×
Pixel dimensions are automatically saved to your browser and restored on reload.
FLUX.2 Klein 4B Distilled Standard: Guidance scale is fixed at 1.0. Diffusers notes step-distilled models bypass guidance scaling.
Optimal 4 steps for distilled Klein checkpoint (~9-12s per image on Vast.ai GPU).
Configurable safety limit for 13–24 GB VRAM. Larger targets (e.g. 3000×3000) generate at 1536px and upscale via Pillow Lanczos filter.
⚡ GPU Backend: Vast.ai
📊 Generation Progress
Standby — no active batch run.
⏳ Generation Queue
0 Songs
| # | Song ID & Title | Target Type | Status | Action | |
|---|---|---|---|---|---|
| Queue is empty. Upload a batch JSON file above to populate the queue. | |||||
🖼️ Completed Artwork (This Batch)
0 Songs Done
No completed artwork yet. Generated images will appear here instantly.
📜 FLUX Engine Event Stream
[00:00:00] [INFO] FLUX.2 Klein image engine initialized in standby mode.
🖼️ Image Artwork Library
0 Songs Completed
No completed artwork found. Completed assets from Image Gen will appear here.