GeneLab | AIクリエイティブ研究

主にComfyUI、最近はMLXを使ったAppleSilicon最適化にハマっている #AICUAward C3AE

[GeneLab | AIクリエイティブ研究 @geneLab_999]

@suntao8023 Thanks—useful control. My case was a 377-character unpunctuated Chinese run in audio.cpp/Magpie, fixed in #1045. Here are the Oct 22 before/after WAVs, exact text and commands:
https://t.co/ZGgwHQXZsw
Different engine; no mid-UTF-8 split observed.

[GeneLab | AIクリエイティブ研究 @geneLab_999]

I've been contributing to audio.cpp, improving long-form multilingual TTS across Ming 3.0, MOSS-TTS-Nano, Kokoro, FireRedTTS3, MagpieTTS and more.

Fixing CJK text truncation and chunking issues, backed by RTX 5090 benchmarks.

PRs:
https://t.co/7qtaY5E8gj

[GeneLab | AIクリエイティブ研究 @geneLab_999]

Apache-2.0, with a JSON CLI and fetch examples for Python and Node.

My project; the code and proofs were written with Claude Code (Claude Opus 5.5) and reviewed only by AI so far, no human review yet. Counterexamples are very welcome.

As of ea3c061: https://t.co/hHnCiNZY4F

[GeneLab | AIクリエイティブ研究 @geneLab_999]

URL host allowlists keep getting it wrong, in AI agents' web tools and elsewhere: startsWith checks (CVE-2026-24052, CVE-2026-46341), a suffix with no dot (CVE-2026-33393, Discourse), userinfo (CVE-2025-47241).

URLFence: the same check, with Lean 4 proofs 🧵

[GeneLab | AIクリエイティブ研究 @geneLab_999]

DNS: urlfence addr checks resolved IPs against IANA's special-purpose registries, incl. mapped, NAT64, 6to4 and Teredo forms. The fetch example checks each redirect and connects only to a checked IP.

Not covered: unpinned clients, IPv6/IDNA URLs, your own hosts on public IPs.

[GeneLab | AIクリエイティブ研究 @geneLab_999]

完売してた、、

[アプライド一宮店 @ap_ichinomiya]

とんでもない衝撃価格....😱

こ、これがアプライドの
リニューアルOPENセールです（震え声）

[GeneLab | AIクリエイティブ研究 @geneLab_999]

@CleanFix_Dev @netadror @Open_MOSS @LexnLin Thanks for checking out ComfyUI-DBS and sharing your test results! Really appreciate you taking the time to try both models and being clear about what was tested.

[GeneLab | AIクリエイティブ研究 @geneLab_999]

こっちも紹介嬉しい！

[CleanFix Dev @CleanFix_Dev]

@netadror @geneLab_999 MOSS-Image-SFT by @Open_MOSS: open text-to-image model, Apache-2.0.

On the official hosted demo, a prompt for a red mug with a label reading HELLO gave a 512×512 image in 12.2 s. Mug and lettering were right.

Catch: one image, hosted only. We did not run it locally.

[GeneLab | AIクリエイティブ研究 @geneLab_999]

紹介うれしい！

[CleanFix Dev @CleanFix_Dev]

@netadror ComfyUI-DBS by @geneLab_999: splits audio into speech, music and effects stems. MIT.

We ran its BSRNN and TIGER models on CPU on a 4-second test tone. Stems summed back with error under 6×10⁻⁸. 23 s and 43 s.

Catch: no real soundtrack, no listening test, no ComfyUI run.

[GeneLab | AIクリエイティブ研究 @geneLab_999]

I just published How I Formally Verified an AI Agent’s Path Check in Lean 4 https://t.co/aG3JRbkYgM

[GeneLab | AIクリエイティブ研究 @geneLab_999]

@PCjank パーツ価格が安いタイミングでメーカー側が仕入れが出来ていた場合、BTOの方が安いパターンがあります。

パーツの価格高騰が続いているため難しいラインですが、、、。

[GeneLab | AIクリエイティブ研究 @geneLab_999]

有名アイドルになっていない。
おかしい。まだなっていない。
そんな中年男性の疑問とダンスを詰め込んだMV『キャンディチューネン』を、国際AIクリエイターアワード2026に応募しました。
中年の晴れ舞台、ぜひ観てください。
https://t.co/ahjzHQrlu1
#AICUaward

[GeneLab | AIクリエイティブ研究 @geneLab_999]

❤️‍🔥

[むらさん(Murasan) @murasametech]

すみません、燃えるそうです🔥

[GeneLab | AIクリエイティブ研究 @geneLab_999]

@suntao8023 Yes please, that’d be really helpful—thanks! Could you share a couple of complete examples and, if possible, the TTS model/tool you used and where the output gets cut off? I’d like to try reproducing the issue in audio.cpp.

[GeneLab | AIクリエイティブ研究 @geneLab_999]

@banr1_ 同じ日本人エンジニアとして、とても刺激になります！
早速手元で動かして、Linux/WSL 2対応とテスト改善のPRを2本送りました（#22・#23）。少しでもお役に立てれば嬉しいです。応援しています！

[GeneLab | AIクリエイティブ研究 @geneLab_999]

@suntao8023 Thanks! That’s exactly the kind of use case I’m hoping to improve. If you have any Chinese or mixed-language examples that get cut off, I’d love to see them—they’d be helpful for regression testing.

[GeneLab | AIクリエイティブ研究 @geneLab_999]

@ko4ro_ 32GBです！AIをぶん回す時にやはり壁を感じます。

[GeneLab | AIクリエイティブ研究 @geneLab_999]

@AiXsatoshi こわい、、