Home

All projects

Trained models, personal projects, and more.

Models & software projects

Irodori Character Voice

Speech synthesis controlled by character images

By extracting visual features with SigLIP or WD Tagger, the system conditions speech generation to adjust voice timbre and tone without reference audio.

An implementation of speech generation conditioned on features of character illustrations, providing PyTorch inference code, model weights, and an interactive web demo.

Approach
Irodori-TTS + SigLIP / WD Tagger
Resources
Model weights, inference code & web demo
  • Text-to-Speech
  • Multimodal
  • PyTorch

Pixel-space Flow Matching image generation

A Flow Matching model that generates illustrations directly in pixel space without latent compression. The 380M-parameter U-JiT architecture supports multiple resolutions around 512px.

  • Image Generation
  • Flow Matching
  • PyTorch

JiT Anime

Pixel-space Flow Matching image generation

Rather than relying on latent compression, this project trained a model to predict raw pixel values on an illustration dataset, using U-JiT to accommodate variable resolutions around 512px.

The repository provides training code alongside an experiment article and a web demo. Note that the published model weights correspond to a separate AnimeFace experiment.

Approach
U-JiT / Pixel-space Flow Matching
Resources
Training code, web demo & experiment notes
  • Image Generation
  • Flow Matching
  • PyTorch

Translating Japanese text to Danbooru tags

A language model that converts natural Japanese descriptions into Danbooru tags for image generation. It was trained on LLM-generated synthetic data and includes a custom ComfyUI node.

  • Language Model
  • Synthetic Data
  • ComfyUI

Danbot

Translating Japanese text to Danbooru tags

To assist prompt creation in tag format, paired datasets of Japanese descriptions and tag sequences were synthesized using an LLM to train a compact language model.

Trained model weights and datasets are published on Hugging Face, accompanied by a custom ComfyUI node for integration into image generation pipelines.

Approach
Language modeling / Synthetic data
Resources
Model weights, dataset & ComfyUI node
  • Language Model
  • Synthetic Data
  • ComfyUI

In-place manga translation browser extension

A browser extension that detects comic dialogue and renders translated text directly over manga pages. It pairs AnimeText YOLO with Gemini 2.5 Flash to account for text layout and typography.

  • Browser Extension
  • WXT
  • Computer Vision

YomiComiX

In-place manga translation browser extension

Built on the WXT framework, the extension couples bounding-box text detection with LLM translation to render translated dialogue within speech balloons, accounting for text layout and typography.

The project was showcased at Generative AI Exhibition vol.4, with presentation slides and technical write-ups published alongside a standalone text-detection code sample.

Approach
WXT / AnimeText YOLO / Gemini 2.5 Flash
Resources
Exhibition materials, development article & text detection example
  • Browser Extension
  • WXT
  • Computer Vision

Personal blog documenting projects and research

A personal technical blog built with Astro. It displays Markdown URLs as card previews, using Cloudflare Workers and Hono to fetch Open Graph metadata.

  • Astro
  • Cloudflare Workers
  • Hono

platform

Personal blog documenting projects and research

A personal blog built as a static site with Astro to document projects and technical research notes.

To display URLs in Markdown as card previews, it implements an API on Cloudflare Workers and Hono to fetch Open Graph metadata, accompanied by a published article explaining the implementation.

Approach
Astro / Cloudflare Workers / Hono
Resources
Blog & link card implementation article
  • Astro
  • Cloudflare Workers
  • Hono

LoRA training for erasing diffusion concepts

A training tool that applies concept erasure to LoRA, removing specific artistic styles or subjects from diffusion models. It also supports emphasizing or swapping concepts via prompt tuning and weights.

  • Diffusion
  • LoRA
  • PyTorch

LECO

LoRA training for erasing diffusion concepts

Adapting prior concept-erasing techniques to LoRA training, this tool suppresses or alters the generation tendencies of specific styles and subjects without updating the entire base model.

The project provides local training scripts, a Google Colab notebook, verification model weights on Hugging Face, and comparison images for testing in tools such as Stable Diffusion WebUI.

Approach
Diffusion models / LoRA / Concept erasure
Resources
Training code, model weights & Colab notebook
  • Diffusion
  • LoRA
  • PyTorch

Web and CLI tool for finding public accounts by username

A tool that checks for public accounts with a given username across multiple services in parallel. It is available as a web app deployed on Cloudflare Workers and as a CLI built with Bun and TypeScript.

  • OSINT
  • Cloudflare Workers
  • TypeScript

Mikuru

Web and CLI tool for finding public accounts by username

The web version, deployed on Cloudflare Workers, is built with Hono and React. It allows users to enter a username and select target services in the browser, displaying search progress and results incrementally.

A CLI built with Bun and TypeScript is also provided, supporting target service filtering, JSON/JSONL output, and an agent skill. Both the web version and the CLI share search logic in a monorepo setup.

Approach
TypeScript, Cloudflare Workers, Bun
Resources
Web App / CLI
  • OSINT
  • Cloudflare Workers
  • TypeScript

Danbooru tag generation and completion model for image prompts

A language model that generates and completes Danbooru tags for image generation prompts. It provides model weights based on Mistral and Mixtral, usable via WebUI and ComfyUI extensions.

  • Language Model
  • Prompt Generation
  • ComfyUI

Dart

Danbooru tag generation and completion model for image prompts

Trained on the context of illustration tagging sequences, this language model predicts and completes subsequent tags from input tags. Version 2 adopts Mistral and Mixtral architectures.

Model weights and tag embeddings are published on Hugging Face. Integrations are provided through a Stable Diffusion WebUI extension and a custom ComfyUI node.

Approach
Mistral / Mixtral / Tag generation
Resources
Model weights, tag embeddings & generation tool integrations
  • Language Model
  • Prompt Generation
  • ComfyUI

CLI tool for inspecting safetensors metadata

A Rust CLI tool for inspecting safetensors model files. It displays parameter counts, tensor shapes, and ModelSpec metadata, supporting remote files on Hugging Face.

  • Rust
  • CLI
  • Model Inspection

safemetadata

CLI tool for inspecting safetensors metadata

In addition to local safetensors files, it inspects tensor configurations and ModelSpec entries from file headers for remote Hugging Face models without downloading entire files.

In addition to displaying total parameter counts and tensor data types, it includes a command to delete metadata from local files. It also supports installation as a CLI skill.

Approach
Rust / safetensors / Hugging Face
Resources
CLI, source code & agent skill
  • Rust
  • CLI
  • Model Inspection