
Cofounder of @fixedseed | https://nitter.cf/t.co/KNLxCQMHT5
Kamakura-shi, Kanagawa
Joined April 2024
- Tweets1.3K
- Following301
- Followers636
- Likes3K
Built my own election tracker in Lattice. It's the coolest IDE I've ever used. It's going public soon! Join the waitlist if you haven't already. We're rolling out bonus perks to beta testers: fixedseed.com/
@fixedseed
カイトラ (Kytra) 凯托拉 retweeted
こんばんは🌙
本日もありがとうございました₍ᐢ..ᐢ₎♡ ༘✩
#SHIBUYA109
#MAGNETbySHIBUYA109
明日も、お近くの【OZZ】&【渋谷2店舗】にて皆様のご来店お待ちしておりますᕱ⑅ᕱ゙
#ozzon #ozzsnap #ozzハロウィン
カイトラ (Kytra) 凯托拉 retweeted
Pika open-sourced Pika H3 Flex: LoRAs on MiniMax H3 that make video + synced stereo audio in 6 steps, with up to 12 refs (9 images, 3 videos, 3 audio clips). Text, first/last frame or ref modes, 480p/768p locally. ComfyUI workflows already up.
huggingface.co/Pika-Labs/Pik…
Biology is crazy man. Did you know a Cricket's chirping is actually a thermometer? youtube.com/shorts/BJl0RjkSJ…
カイトラ (Kytra) 凯托拉 retweeted
2 months after Fal said they're release H3 Max, it remains unreleased
Meanwhile, Pika took the opportunity to release a H3 fine-tune that performs better than it:
カイトラ (Kytra) 凯托拉 retweeted
「Tagger 1.0」をオープンソース化しました
Tsubaki.3の学習データ整備に使用した、アニメ概念タガー「Tagger 1.0」を公開しました。
・骨格:SAM 3のViTDet系画像エンコーダを採用。入力解像度は1008×1008、全32ブロック。高解像度の特徴量を活用し、類似したアニメ関連の概念も細部まで捉えて区別します。
・分類ヘッド:30,877種類のタグを予測
・学習時のみの工夫:小規模な再構成デコーダで特徴量を共有し、視覚的な細部の情報が失われるのを防ぎます。
重み・モデルカード:
huggingface.co/pixai-labs/pi…
【#PixAITsubakiV3 は何ができる?】
PixAIの技術レポートをもとに、Tsubaki.3の能力を詳しく紹介します!
画像生成技術はすでに大きく進歩していますが、アニメイラストの分野では、多様なスタイル、複数のキャラクター、複雑なポーズ、アーティスティックなカメラアングルなど、まだ十分にカバーしきれていないニーズが数多くあります。
Tsubaki.3は、こうした課題に対応するために開発されたマルチモーダル基盤モデルです。テキストからの画像生成とコンテキスト内生成の両方に対応し、アニメ創作コミュニティのために設計されています。
Added LoRA training for Qwen 3.8 27B in AI Toolkit and spent the morning optimizing the captioner for it. It now captions at 70-80 tok/s with 1024 res image inputs and a decently long system prompt on a downclocked 5090 using less than 20GB of vram.
Do the fans that help the Roadster stick to the track ever get hair and sh*t in them or what?
youtube.com/watch?v=wMFfRaSM…
I’m at the ElevenLabs LA Tech Week after party and did you know that ElevenLabs V4 can’t say the word “embedding” in a prompt by itself? It’s says “dot” or begins to say a random url. “http colon dot”
Just something I discovered while having a midnight scrape.
I saw a first hand demonstration of Jackson using Torii and it was sick. I’ve got to explore Telegram automation more.
introducing torii: your agents, your subscriptions, your machine.
grok bot and dots are built on the same idea: with today's frontier models, you can hand work to an agent manager instead of babysitting each agent yourself.
but they all want you locked into their ecosystem. so i built the same thing, on your own computer.
i've been running all my development through one telegram chat for months.
claude code and codex agents, each project in its own thread, a manager handing out work, checks running, PRs opening while i'm away from my desk.
torii is as lean as possible. it does a few things:
- connects claude code or codex to a telegram chat
- tells each agent to prioritize delegation
- helps rotate accounts when a subscription is running low
open source: github.com/pnkfluffy/torii
カイトラ (Kytra) 凯托拉 retweeted
Fish Audio Drama 3: TTS you direct in plain words — tears, whispers, laughs on cue.
Open weights. 83 languages. Tops Artificial Analysis pronunciation robustness.
Blog: fish.audio/blog/fish-audio-d…
Weights: huggingface.co/fishaudio/dra…
カイトラ (Kytra) 凯托拉 retweeted
Implementing full VRM support into Lattice!
Join the waitlist here fixedseed.com/lattice
カイトラ (Kytra) 凯托拉 retweeted
I rebuilt photoshop in type script, you can try it for free here!
We built agentic photoshop in the web!
Try it here for free fixedseed.com/playground/ima…




















