kizuna-intelligence ORG

Text prompt-based voice quality designer using FlowMatching + GPT-SoVITS

Preview-first voice studio for Kizuna Voice Designer, Qwen Voice Designer, MioTTS, Piper, and Style-Bert-VITS2

Int4-quantized inference runtime for Faster-Irodori-TTS2 voice-design DiT. ~1GB VRAM end-to-end.

Inference runtime that runs the Irodori Japanese DiT TTS model with GPTQ 4-bit quantization, shrinking VRAM use from 1.88GB to about 552MB with preserved audio quality.

Int4-quantized FLUX.2-klein inference runtime (GemLite / fused Triton / eager backends)

A Python library that compresses the FLUX.2-klein image generation model using 4-bit quantization so it runs on GPUs with as little as 2.7 GB of memory instead of the usual 10 GB.

No description.

Swift package for Piper-based on-device TTS on iOS with swappable model integration and demo app

Run the 218B Command-A-Plus MoE on a single 24GB GPU — int2 CPU-offloaded experts + int4 resident weights.

Python package for LLM compression

Shared int4 inference runtime extracted from OneCompression per-model runtimes (GPTQ/RTN packed-int4 GEMM kernels + generic diffusion loader).

A Flow Matching-based Text-to-Speech Model with Emoji-driven Style Control

High-Quality Voice Cloning TTS for 600+ Languages

GitHub organization profile for Kizuna Intelligence