Generative AI, in Japanese, on your own machine.

AI engineer and technical consultant in Tokyo. I build local AI tools, train small Japanese language models, and turn research implementations into usable software for Windows, Linux and Apple Silicon. My ComfyUI integrations are linked from the official READMEs of LongLive-Plug (NVlabs), SparkDiffusion (AlibabaResearch), LeapTalk, Causal Forcing and Looped-DiT.

日本語 · 简体中文 · Hugging Face · Zenn · X

Looking Ahead integrates the 1.3B Future Chunk Prediction model with ComfyUI's existing umT5-XXL and Wan VAE, standard VIDEO / IMAGE / LATENT outputs, cancellation and model unloading. Tested on Windows / RTX 5090 at the default 768×432, 16 fps, five seconds. The sampler matches upstream bit for bit with aligned inputs; end-to-end video outputs differ.

Windows inference fixes were also merged upstream: SDPA fallback when FlashAttention 4 cannot import, and direct model loading without fcntl.

11 PRs merged into audio.cpp as of October 24, 2026. Improved long-text handling across seven TTS families—Ming 3.0, Kokoro, MOSS-TTS-Nano, Irodori, FireRedTTS3, MagpieTTS and MioTTS—to reduce dropped or garbled speech in Japanese, Chinese and Korean. Also fixed shared chunking for punctuation-free CJK text. Ming now receives CLI/server language settings and preserves the designed voice across long-text chunks.

The first Server, Doctor and Text to Speech nodes for ComfyUI-AudioCPP were also merged upstream. They connect ComfyUI to a local audio.cpp server and return standard AUDIO, with optional reference audio for voice cloning.

• XiaomiLens: text-driven instance segmentation with full-resolution masks and boxes. My SDPA fallback fix was merged upstream, reducing mask-decoder memory use when flash-attn is unavailable.
• DBS · VoxLock · MimeticA0: separate music, effects and dialogue; isolate a target speaker; generate 48 kHz stereo music from text.
• S2Tok · EmbeddingGemma2 · SyncReward: reconstruct static scenes as 3D Gaussians, compare text/image/video/audio embeddings, and score audio–video alignment.

A 314.6M-parameter ModernBERT-ja model that returns typed choice, score and bool answers with probabilities in one forward pass. Includes an HTTP server, a TypeScript SDK and a CLI; inference uses MLX on supported Apple Silicon Macs or PyTorch on CUDA / CPU.

v0.3.0 is available through five channels: PyPI · npm · Homebrew tap · Scoop bucket · GHCR.

The npm package combines a CLI wrapper and a typed HTTP SDK. Homebrew and Scoop install launchers; the container targets Linux amd64 CPU. CI covers installation, SDK/server integration and container model inference. Installation guide · Model weights

An unofficial integration of JoyAI-Video-Edit, with a persistent WSL2 worker, model load/unload controls, cancellation and memory guards. v0.1.0 pre-release, tested on Windows 11 with an RTX 5090: 840×480 landscape output, two inference steps, up to 600 input frames.

Validation compares against a reference with the same low-memory placement patches; an unmodified upstream end-to-end run was resource-blocked. Production outputs are not bitwise repeatable. Setup and validation scope

I publish code, setup instructions and measured results together. Hardware, inputs, upstream differences and untested paths are documented in each project's README. ComfyUI integrations are community projects; upstream code and weights retain their own licenses.

1LM → 2LM → 3LM explores character tokenization, subwords and modern decoder architectures on a single machine. Base models are trained from scratch; GAL variants study the trade-off between persona fine-tuning and general ability.

21 node packs published on the ComfyUI Registry as of October 20, 2026, under hiroki-abe-58. Selected projects below cover generation, editing, analysis and 3D reconstruction.

The Upstream column distinguishes merged code fixes, README listings and the projects an integration is based on.

Languages & scripting

AI & model runtimes

Web & desktop

Build, delivery & platforms

Built across my ML projects, React libraries, desktop apps, web tooling and Go CLIs. Docker Compose is also part of my everyday workflow. Main hardware: RTX 5090 (Blackwell) and Apple Silicon. Badges without a brand logo use a generic code icon.

Looking Ahead · Windows inference

#2 fixes FlashAttention 4 import failures, routes cross-attention through the SDPA fallback and loads models directly when fcntl is unavailable. The upstream 1.3B example completed on Windows 11 / RTX 5090: 80 frames at 768×432, with 16.4 GiB peak allocated memory. The existing FA4 path retained bit-identical decoded frames in the comparison test; SDPA outputs differ numerically.

audio.cpp · 11 merged PRs (October 24, 2026)

• Unpunctuated CJK text: #1045 splits over-budget runs of hanzi, kana and Hangul into balanced chunks, preventing punctuation-free ASR/OCR text from bypassing the chunk limit. Verified on MagpieTTS, FireRedTTS3 and MOSS-TTS-Nano; punctuated regression cases stayed byte-identical.
• Long-text synthesis across seven model families: language-aware defaults for Ming #1020, Kokoro #1022, MOSS-TTS-Nano #1025, Irodori #1027, FireRedTTS3 #1039, MagpieTTS #1043 / #1044 and MioTTS #1046. Magpie's follow-up uses 100 codepoints to avoid a mixed Chinese/English regression. Validation includes Windows / RTX 5090 synthesis, Whisper CER measurements and output comparisons where reproducible.
• Ming language and voice consistency: #1021 forwards CLI/server language settings; #1034 splits voice-design/freeform text and reuses the first generated segment as the voice reference. On a 404-character Japanese voice-design test, mean CER fell from 0.913 to 0.043 over four seeds on the tested setup.
• ComfyUI integration (separate repository): ComfyUI-AudioCPP #2 adds Server, Doctor and Text to Speech nodes. Validated through 21 real ComfyUI workflows; successful generated audio matched direct server requests sample for sample.

Inference fixes merged upstream

• Xiaomi-Lens-0: Use SDPA when flash-attn is unavailable, avoiding quadratic attention-weight allocations in the mask decoder. On Windows / RTX 5090, added peak VRAM for a 3 MP mask pass fell from 18.02 GB to 2.28 GB; tokens and boxes stayed the same, with small mask-edge differences.
• SemBridge: Remove TorchScript from the codec's Snake activation to fix audio changing between the first generation after loading and later calls with the same request and seed. Verified repeated outputs on the tested setup.

LLM inference & reproducibility

• Project Maya: Documented how to reproduce greedy outputs across restarts with the CPU lane enabled (merged). Investigated CPU/GPU rounding differences and expert placement, and validated settings that reproduced all 70/70 token-stream pairs on an RTX 5090. Experiments and results.

Image and video generation

• DMAD (distillation for fast image and video generation): made --low-vram work on Windows, merged and credited in the README News.
• SparkDiffusion, LongLive, LeapTalk and Looped-DiT: README entries for my ComfyUI integrations, merged (see the ComfyUI table above).
• Mi-Ripple: OpenCV runtime handling (open), found while validating the ComfyUI integration.

Decision models

Reproducible checks for answer-order sensitivity:

• Laya: score-position bias report, regression checks, non-English inputs and choice questions.
• kev / lev: kev measurements, lev measurements and presentation checks / order averaging.

DiM-2 explores Mamba-2 / Structured State Space Duality for image and video diffusion. These preprints present architectural proposals and experimental directions; implementation results are reported separately in the OSS repositories above.

I write about implementation and experiments in Japanese and English.

Zenn · Qiita · note · Medium · dev.to · X

If these tools save you time, a coffee helps me maintain them.