🚀 IvanderLab Colab AI Suite

High-Performance, Zero-Token Google Colab GPU Deployments for Generative AI

A unified collection of production-ready, zero-token Google Colab notebooks curated and engineered by IvanderLab. Designed specifically for creators, voice actors, and developers who want to run state-of-the-art Generative AI models—including SDXL/Pony image generation, low-latency realtime voice conversion, emotional few-shot TTS, AI song covers, and local narrative LLMs—on Google Colab's Free Tier (NVIDIA Tesla T4 GPU).

---

⚡ Quick Access Catalog

Click any Open in Colab badge to launch the environment immediately in your browser:

Notebook	Domain	Core Engine / Model	Hardware Profile	Launch
🎨 SD-WebUI Forge (Pony XL)	Image Generation	SD-WebUI Forge • AutismMix SDXL • 18 LoRAs	Tesla T4 (FP16)
🎙️ W-Okada Realtime Voice Changer	Realtime Audio	MMVCServerSIO • RVC v1/v2 • ContentVec	Tesla T4 (~67ms)
🗣️ GPT-SoVITS Emotional TTS	Voice Synthesis	GPT-SoVITS • Zero-Shot Voice Cloning	Tesla T4
🎵 Applio RVC Voice Studio	Cover & Training	Applio RVC • Edge-TTS • Audio Separator	Tesla T4
🧠 KoboldCpp x L3.1 Stheno 8B	Narrative LLM	KoboldCpp • L3.1-8B-Stheno-v3.2 (Q5_K_M)	Tesla T4 (Offload)

---

🎨 1. SD-WebUI Forge (Pony XL Edition)

File: Ivan_Forge_Pony_Masterpiece.ipynb
Key Capabilities: AutismMix SDXL, 18 Curated Anime LoRAs with Visual Previews, ADetailer Facial Mesh, Danbooru TagComplete, Infinite Image Browser, Stealth Anti-Ban Launch System.

Overview

Stable Diffusion WebUI Forge is an optimized fork of AUTOMATIC1111 providing up to 2x faster generation and 50% less VRAM allocation. This suite delivers an out-of-the-box anime and illustration environment tuned for Pony Diffusion V6 XL and AutismMix SDXL. It bypasses Colab's automated static code scanners through an encrypted repository unban layer.

Step-by-Step Walkthrough

1. CELL 1: Setup Environment & Stealth Unban
   Prepares Python dependencies, pulls the optimized Forge core from camenduru/forge, and installs essential UI extensions:
   • sd-webui-infinite-image-browsing
   • adetailer (automated face and hand restoration)
   • a1111-sd-webui-tagcomplete (Danbooru auto-tag completion)
2. CELL 2: High-Speed Model Downloader
   Downloads checkpoint weights via multi-threaded aria2c at ~80–120 MB/s:
   • Base Checkpoint: AutismMix SDXL (autismmixSDXL_autismmixConfetti.safetensors)
   • VAE: Native SDXL VAE (sdxl_vae.safetensors)
   • 18 Anime Style & Character LoRAs: Gacha styles, dynamic lighting, watercolor, cyber-aesthetic, and expression modifiers.
3. CELL 3: Launch Stealth Forge WebUI
   Initializes Forge with optimal GPU flags (--opt-sdp-attention --enable-insecure-extension-access) and provisions a public Cloudflare tunnel.

Recommended Generation Parameters

• Sampler: Euler a or DPM++ 2M Karras
• Sampling Steps: 28 – 35
• CFG Scale: 5.5 – 7.0 (Do not exceed 8.0 on Pony models to prevent chromatic aberration)
• Resolution: 832 x 1216 (Portrait) or 1024 x 1024 (Square)
• Mandatory Positive Prefix: score_9, score_8_up, score_7_up, source_anime, rating_safe, ...
• Mandatory Negative: score_6, score_5, score_4, simple background, bad anatomy, deformed limbs

---

🎙️ 2. W-Okada Realtime Voice Changer (RVC v1 & v2)

File: Colab_RealtimeVoiceChanger.ipynb
Key Capabilities: Native CUDA Inference, 67ms GPU Latency, TorchScript JIT Fix, Auto-Dimensional Calibration, Cloudflare HTTP/2 Protocol.

Overview

A heavily patched deployment of w-okada / tg-develop MMVCServerSIO engineered to eliminate common tensor mismatch runtime errors on PyTorch 2.x and Google Colab Free Tier.

Engineering Breakthroughs

• TorchScript JIT Subclass Tensor Fix: Solves the upstream exception where RVC v2 (768-dim) models inherit from RVC v1 (256-dim) base classes, preventing RuntimeError: Expected size for first two dimensions of batch2 tensor to be: [1, 256] but got: [1, 768]. Our configuration forces "disableJit": 1 in native CUDA eager execution.
• Auto-Calibration Pipeline: Automatically inspects checkpoint weights and configures embChannels (256 vs 768), embOutputLayer (Layer 9 vs Layer 12), and sampling rate (32k/40k/48k).
• HTTP/2 Cloudflare Tunneling: Mitigates Indonesian & Southeast Asian transpacific ISP UDP packet-drop by enforcing --protocol http2.

Telemetry & Benchmarks (499 Realtime Frames)

Metric	Mean	Median (50th)	Min	95th Percentile	Status
GPU Inference (PerfValue)	67.72 ms	65.00 ms	59.00 ms	83.00 ms	99.4% Good
Audio Chunk Size (PerfChunk)	162.06 ms	168.00 ms	131.00 ms	168.00 ms	Stable
Network Roundtrip (Ping)	641.20 ms	641.00 ms	314.00 ms	1,181.00 ms	Network Bound
Total End-to-End Latency	907.40 ms	907.00 ms	545.00 ms	1,371.00 ms	Playable

Full raw dataset: benchmarks/performance_telemetry.csv.

Recommended WebUI Settings

• GPU ID: cuda:0 (Tesla T4)
• Chunk Size: 170.7 ms – 341.3 ms (absorbs ~500ms network jitter without robotic stutter)
• Extra Processing Time: 0.5 s (4096 samples)
• Pitch Detection (f0): rmvpe
• FAISS Index Ratio: 0.6 – 0.7
• Input Gate: -55 dB

---

🗣️ 3. GPT-SoVITS (Zero / Few-Shot Emotional TTS)

File: Colab_GPT_SoVITS_TTS.ipynb
Key Capabilities: 5-Second Voice Cloning, Multilingual Synthesis (Indonesian, English, Japanese, Chinese), Emotional Intonation Control.

Overview

Unlike RVC which requires a live microphone input to modulate pitch, GPT-SoVITS is a state-of-the-art Text-to-Speech system. It clones any character or human voice from a single short audio sample (3 to 10 seconds) and synthesizes completely new spoken sentences with authentic cadence and emotion.

Step-by-Step Walkthrough

1. CELL 1: Setup Environment & Base Weights
   Clones the repository, installs FFmpeg and PyTorch audio dependencies, and automatically retrieves base pretrained Chinese/English/Japanese HuBERT models.
2. CELL 2: Model & Sample Ingestion
   Downloads pre-indexed character voice packages directly into the cache directory.
3. CELL 3: Launch TTS WebUI
   Generates a Gradio and Cloudflare live tunnel.
4. Usage in WebUI:
   • Upload or select a 5-second clean reference audio clip.
   • Provide the transcript of what was said in the reference clip.
   • Type any target text in the prompt box and click Synthesize Speech.

---

🎵 4. Applio RVC Voice Studio

File: Applio.ipynb
Key Capabilities: Edge-TTS Script Voiceover, AI Song Cover Generation, UVR5 Vocal & Instrumental Demixing, Offline Audio Processing.

Overview

Applio (by IAHispano) is an advanced modular RVC ecosystem optimized for batch processing, automated voice generation, and high-fidelity vocal track conversion.

Core Features

• AI Song Cover Pipeline: Input any MP3/WAV song; Applio automatically splits vocals and instrumental stems using UVR5 algorithms, shifts pitch, replaces the singer's voice with your chosen RVC model, and remixes the final track.
• Edge-TTS Narrator: Convert lengthy markdown or article scripts into spoken natural narration using Microsoft Edge neural voices (e.g., id-ID-GadisNeural, id-ID-ArdiNeural), and instantly convert them into anime character voices in one click.

---

🧠 5. KoboldCpp x L3.1 Stheno 8B (Narrative Engine)

File: Stheno_KoboldCpp_Colab.ipynb
Key Capabilities: Llama 3.1 8B Architecture, Sao10K Stheno v3.2 Checkpoint, KoboldCpp C++ Engine, Native RisuAI & SillyTavern Tunnel.

Overview

Runs L3.1-8B-Stheno-v3.2—one of the highest-rated open-weights narrative, creative writing, and roleplay models—on Google Colab Free Tier using the hyper-optimized C++ inference engine KoboldCpp.

Features & Setup

1. Anti-Sleep Keep-Alive: Cell 1 plays a silent looping HTML5 audio buffer to prevent Chrome and Edge tabs from being suspended by Google Colab while working in another window.
2. KoboldCpp Server: Loads the Q5_K_M GGUF quantization fully into Tesla T4 VRAM with GPU offload (-ngl 99), delivering inference speeds of ~45+ tokens per second.
3. Integration with Frontends:
   • Copy the generated https://xxxx.trycloudflare.com URL.
   • In RisuAI or SillyTavern, set API to KoboldCpp / Text Generation WebUI and paste the URL.
   • Enjoy full local context window processing without token limits or API subscriptions.

---

📁 Repository Structure

IvanderLab-Colab-Suite/
├── assets/
│   ├── hero_banner.svg                  # Modern vector hero header banner
│   ├── vision_showcase.png              # 3-grid high-resolution vision showcase
│   └── audio_pipeline.svg               # Vector architecture flowchart
├── Ivan_Forge_Pony_Masterpiece.ipynb    # SDXL & Pony V6 Stable Diffusion Forge Studio
├── Colab_RealtimeVoiceChanger.ipynb     # W-Okada Realtime Voice Changer (Patched RVC)
├── Colab_GPT_SoVITS_TTS.ipynb           # GPT-SoVITS Zero-Shot Emotional TTS
├── Applio.ipynb                         # Applio Song Cover & Edge-TTS Suite
├── Stheno_KoboldCpp_Colab.ipynb         # KoboldCpp Llama 3.1 8B Stheno LLM Engine
├── benchmarks/
│   └── performance_telemetry.csv        # 499-frame raw latency & GPU telemetry log
├── LICENSE                              # MIT License
└── README.md                            # Suite documentation & walkthroughs

---

💡 Best Practices for Free-Tier Colab Users

1. Avoid Multiple Concurrent Sessions: Google Colab Free limits each Google account to one active GPU instance at a time. Always disconnect your runtime before opening another notebook.
2. Keep Browser Active: While running realtime audio or LLM generation, ensure the browser tab remains active or run the embedded silent audio keep-alive cell.
3. Zero Token Philosophy: All notebooks in this suite use zero-token public tunnels (Cloudflare Tunnel / trycloudflare.com). No account registrations, credit cards, or API secrets are required.

---

👨‍💻 Author & Maintainer

Andhika Ivander Nathanael (Nat / Ivander)

• Brand: IvanderLab (@ivanderlab)
• Role: Creative Developer & DKV / Interactive Media Designer
• Contact: contact.ivanderlab@gmail.com
• Socials: GitHub • Instagram • LinkedIn • Behance

---

📜 License

This project is licensed under the MIT License.