Lora beYond Conventional methods. Comprehensive PEFT library for neural networks, published at ICLR 2024.

Text to Image with Text Presampling for Prompt Optimization. Published at ICLR 2026, related models got over 3M downloads.

Home-made Diffusion Model. Training quality text-to-image models on consumer-level hardware. Tech report on arXiv.

General purpose AI agent framework with built-in tools, sub-agents, persistent sessions, TUI and web UI. Comes with kt CLI and package marketplace.

AI accelerator implementation with Xilinx FPGA. Built on KohakuAccel, an open platform for building FPGA accelerators.

Detail-Oriented Pixelization based on Contrast-Aware Outline Expansion.

Self-hosted HuggingFace alternative, fully compatible with HuggingFace Hub, Transformers and Diffusers.

#1 in Wattbot2025. Hierarchical RAG framework with context-aware search and single-file SQLite storage. Paper on arXiv.

Lightweight cluster manager that turns a small fleet of nodes into one powerful computer using Docker. Supports QEMU/KVM VM, VXLAN overlay network and GPU passthrough.

KIR: a text-based intermediate representation for visual programming. Captures control flow and dataflow in one file, round-trips between text and graph, runs on any conforming backend.

Unified black-box optimization system for hyperparam tuning. Reference implementation of S3-TuRBO with from-scratch baselines (CMA-ES, GP-BO, HEBO, TuRBO).

Single-file key-value/columnar store built on SQLite with Rust. Supports full-text search, vector search and auto msgpack.

Self-hosted W&B alternative with async logging, efficient storage and fast visualization frontend.

A comprehensive codebase for training and finetuning Image <> Latent models. Supports EQ-VAE regularization.

Board game engine with NNUE neural network evaluation and PVS search. Supports MiniChess, MiniShogi, and Gomoku.

KohakuLayout: netlist to placed-and-routed grid layout framework. KohakuEFDA: factory planner for Arknights: Endfield built on it.

Triton kernels for Key-Value Means (KVM) attention with hand-written backward. Also a scheduler playground for monotone bucket schedules.

Unified Perceptual Feature Loss Framework. LPIPS, ConvNeXt/DINOv3 feature loss, frequency domain loss and PSNR/SSIM metrics.

All-in-Python config and execution engine for RD workload. Run existing scripts with different configs without modifying the code.

Stateless Rust+pyo3 video loader tailored for fast, memory-efficient PyTorch pipelines.

Simple, fast and robust async S3/HTTP downloader and uploader with pipelined parallel transfers. Can patch huggingface_hub for faster transfer.

A sd-webui/ComfyUI extension for utilizing TIPO or DanTagGen to "upsample prompts".

An extension for stable-diffusion-webui to load LyCORIS models.

text-image dataset maker for anime-style images.

A Novel-AI client with more utilities built in it.