https://kotonia.ai/ 2026-05-18T14:10:42+09:00 weekly 1.0 https://kotonia.ai/en/ 2026-05-18T14:10:42+09:00 weekly 0.9 https://kotonia.ai/zh/ 2026-05-18T14:10:42+09:00 weekly 0.9 https://kotonia.ai/studio/ 2026-05-28T23:40:00+09:00 weekly 0.9 https://kotonia.ai/en/studio/ 2026-05-28T23:40:00+09:00 weekly 0.8 https://kotonia.ai/zh/studio/ 2026-05-28T23:40:00+09:00 weekly 0.8 https://kotonia.ai/voice-studio/ 2026-05-26T12:31:41+09:00 weekly 0.8 https://kotonia.ai/en/voice-studio/ 2026-05-26T12:31:41+09:00 weekly 0.7 https://kotonia.ai/zh/voice-studio/ 2026-05-26T12:31:41+09:00 weekly 0.7 https://kotonia.ai/video-studio/ 2026-05-18T14:10:42+09:00 weekly 0.8 https://kotonia.ai/en/video-studio/ 2026-05-18T14:10:42+09:00 weekly 0.7 https://kotonia.ai/zh/video-studio/ 2026-05-18T14:10:42+09:00 weekly 0.7 https://kotonia.ai/articles/ 2026-05-16T02:06:13+09:00 daily 0.9 https://kotonia.ai/en/articles/ 2026-05-16T02:06:13+09:00 daily 0.8 https://kotonia.ai/zh/articles/ 2026-05-16T02:06:13+09:00 daily 0.8 https://kotonia.ai/about/ 2026-05-19T18:31:17+09:00 monthly 0.5 https://kotonia.ai/en/about/ 2026-05-19T18:31:17+09:00 monthly 0.4 https://kotonia.ai/zh/about/ 2026-05-19T18:31:17+09:00 monthly 0.4 https://kotonia.ai/how-to-use/ 2026-05-21T06:29:02+09:00 weekly 0.8 https://kotonia.ai/en/how-to-use/ 2026-05-21T06:29:02+09:00 weekly 0.7 https://kotonia.ai/zh/how-to-use/ 2026-05-21T06:29:02+09:00 weekly 0.7 https://kotonia.ai/why/ 2026-05-25T12:35:35+09:00 weekly 0.85 https://kotonia.ai/en/why/ 2026-05-25T12:35:35+09:00 weekly 0.75 https://kotonia.ai/zh/why/ 2026-05-25T12:35:35+09:00 weekly 0.75 https://kotonia.ai/use/ai-chat/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/ai-chat/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/ai-chat/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/voice/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/voice/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/voice/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/personas/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/personas/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/personas/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/multi-model/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/multi-model/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/multi-model/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/debate/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/debate/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/debate/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/agent/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/agent/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/agent/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/agent-runtime/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/agent-runtime/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/agent-runtime/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/tools/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/tools/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/tools/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/group/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/group/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/group/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/wiki/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/wiki/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/wiki/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/booking-bot/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/booking-bot/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/booking-bot/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/group-booking/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/group-booking/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/group-booking/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/use/outbound/ 2026-05-24T07:44:22+09:00 weekly 0.7 https://kotonia.ai/en/use/outbound/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/zh/use/outbound/ 2026-05-24T07:44:22+09:00 weekly 0.6 https://kotonia.ai/pricing/ 2026-05-22T16:45:08+09:00 weekly 0.8 https://kotonia.ai/legal/tokusho/ 2026-05-08T05:55:46+09:00 monthly 0.3 https://kotonia.ai/use/mesugaki-english/ 2026-05-22T11:42:44+09:00 weekly 0.7 https://kotonia.ai/zh/articles/ltx2-cold-start-vram-coexistence/ 2026-05-28T15:21:24.254059Z https://kotonia.ai/articles/ltx2-cold-start-vram-coexistence/cover.png LTX-2.3 在单张 96GB GPU 上与 TTS 共存的冷启动方案 在语音角色扮演产品中集成 LTX-2.3 时，持久化模式会占用 86 GiB VRAM，无法与 TTS/Ditto 共存。本文记录了切换到冷启动模式，实现空闲 0 GiB / 峰值 40 GiB 的实施方案。 weekly 0.6 https://kotonia.ai/zh/articles/ltx2-22b-fp8-cast-quantization/ 2026-05-28T15:21:24.283928Z https://kotonia.ai/articles/ltx2-22b-fp8-cast-quantization/cover.png LTX-2 22B 使用 fp8_cast 将峰值显存降低 40% — optimum-quanto 是个陷阱 记录 LTX-2.3 22B 的量化尝试。optimum-quanto 与 LTX-2 transformer 存在兼容性问题无法运行，改用 LTX-2 原生的 QuantizationPolicy.fp8_cast() 后，峰值显存从 40 GiB 压缩至 24 GiB（cold-start, 768×512）。包含 3 种分辨率的基准测试以及 cold-start / persistent 模式的选择判断。 weekly 0.6 https://kotonia.ai/zh/articles/comedy-shorts-claude-gemini/ 2026-05-28T15:21:24.320903Z https://kotonia.ai/articles/comedy-shorts-claude-gemini/cover.png 用 Claude Code 复刻语言学习喜剧短视频——将 Gemini 作为子代理的多模态扩展实践 本文介绍了如何在 Claude Code 开发环境中，通过本地 GPU + Gemini 3.1 Pro Preview 混合架构，从一行创意文本生成 Pingo 风格的语言学习喜剧短视频。重点探讨了将前沿模型作为子代理，在不膨胀主代理上下文的前提下获取高质量编辑信号的方法。 weekly 0.6 https://kotonia.ai/zh/articles/hidream-skeleton-pose-prompt/ 2026-05-28T15:11:55.431026Z https://kotonia.ai/articles/hidream-skeleton-pose-prompt/A_baseline.jpg HiDream skeleton：OpenPose 参考图不如 prompt 强大（8 种模式实证） 对 HiDream-O1-Image (8B Full) 的 skeleton 模式进行 8 种模式 + 3 种布局模式的实证基准测试。通过解读 pipeline.py，解释为何传入 openpose 参考图反而会导致姿势固化，以及通过 prompt 指定为何更强大。 weekly 0.6 https://kotonia.ai/zh/articles/hidream-quality-speed-bench/ 2026-05-28T15:03:19.054034Z https://kotonia.ai/articles/hidream-quality-speed-bench/presets.jpg HiDream-O1-Image 提速 3~8 倍：steps / CFG / 分辨率实测基准 实测 HiDream-O1-Image Full 在不同 steps、CFG 和分辨率下的生成速度与画质，找到兼顾效率与质量的实用预设，让探索迭代成本大幅降低。 weekly 0.6 https://kotonia.ai/zh/articles/96gb-vram-mma-and-5-years/ 2026-05-28T15:02:50.755402Z https://kotonia.ai/articles/96gb-vram-mma-and-5-years/case_B_storyboard_local_llm.png 从3万日元的Chromebook到96GB GPU——五年不放弃的故事 一位个人开发者从3万日元的Chromebook起步，历经五年，最终购入96GB VRAM的RTX PRO 6000 Blackwell GPU。本文记录了他从Colab凑合、AI创业受挫、到自购GPU的历程，以及96GB如何改变他的开发能力。 weekly 0.6 https://kotonia.ai/zh/articles/consent-dilemma-10beat-pipeline/ 2026-05-28T15:21:24.356657Z https://kotonia.ai/articles/consent-dilemma-10beat-pipeline/cover.png 一行创意到40秒喜剧视频：10-beat全自动流水线 面向本地GPU开发者的AI视频喜剧量产方案。记录从一行创意，经Gemma 4、HiDream、LTX-2等模型，全自动生成40秒竖屏视频（含Hook和字幕）的10-beat流水线设计、陷阱与代码结构。 weekly 0.6 https://kotonia.ai/zh/articles/mesugaki-english-persona-gemini-audio/ 2026-05-28T15:21:24.399509Z https://kotonia.ai/articles/mesugaki-english-persona-gemini-audio/cover.png 小恶魔英语会话拿下搜索第一，然后失去的故事 个人开发的 niche AI 英语会话“小恶魔英语学习”为何能戳中用户，拿下搜索第1、2位的策略，以及自家原创内容被 Zenn 夺走的始末。 weekly 0.6 https://kotonia.ai/zh/articles/hidream-prompt-enhancer/ 2026-05-28T15:01:22.794558Z https://kotonia.ai/articles/hidream-prompt-enhancer/enhanced-anime.jpg HiDream 的“随手出图”崩了 → 转投 Dev → VRAM 爆表 → 用提示词工程打赢的实战记录 HiDream-O1-Image 在日语提示词下“随手出图”崩坏，本想换 Dev-2604 却因 VRAM 不足和编辑性能下降而放弃，最终用 Gemini Flash Lite 打造的提示词增强器解决。文中还发现了 HiDream 特有的 4 个陷阱（品牌名嵌入、cute 导致幼体体型、王家卫触发韩文字幕、idol-class 自动加字幕），并通过 A/B 测试验证。 weekly 0.6 https://kotonia.ai/zh/articles/residual-video-domain-i2v/ 2026-05-28T15:21:24.433668Z https://kotonia.ai/articles/residual-video-domain-i2v/cover.png “大厂退出的视频领域”，个人实验全记录——从模型对比到以 I2V 为核心 去掉安全限制的自由创作是否存在小众需求？本文记录了个人开发者用单张本地 GPU，从模型 A/B 测试到将高分辨率 I2V 作为主力的试错过程。 weekly 0.6 https://kotonia.ai/zh/articles/zenn-canonical-crosspost-seo/ 2026-05-28T15:21:24.462736Z https://kotonia.ai/articles/zenn-canonical-crosspost-seo/cover.png Zenn 的 canonical 我以为写了，其实根本没生效 个人开发服务的注册数下降，重新调查流量来源和 canonical 后发现，向 Zenn 全文交叉发布的运营策略存在漏洞。 weekly 0.6 https://kotonia.ai/zh/articles/hidream-o1-lora-why/ 2026-05-28T14:59:21.532095Z https://kotonia.ai/articles/hidream-o1-lora-why/showcase2500_04_off_vs_on.png HiDream-O1-Image 自制 LoRA 的理由——不藏私版 详解为 HiDream-O1-Image 制作通用动漫/半写实质感增强 LoRA 的动机、191 张手动数据收集、NSFW 转换策略，以及 noindex 文章竟成唯一浏览榜首的故事。技术细节的日语版。 weekly 0.6 https://kotonia.ai/articles/features-existed-ux-was-the-bottleneck/ 2026-05-28T14:38:49.696120Z https://kotonia.ai/articles/features-existed-ux-was-the-bottleneck/hero.png 個人開発で機能はあるのに使われてない問題、UI 分割で解いた 動画生成経由で登録した新規ユーザーが初日に離脱する問題を行動ログから分析。ReAct エージェントに既に存在した動画生成ワークフローを、フロントの独立サーフェスとして開放した話。 weekly 0.7 https://kotonia.ai/en/articles/features-existed-ux-was-the-bottleneck/ 2026-05-28T14:38:49.696120Z https://kotonia.ai/articles/features-existed-ux-was-the-bottleneck/hero.png When the Features Were There All Along — How a UI Split Saved My Solo SaaS's Onboarding A first-party analytics post-mortem: video generation drives signups but day-one churn was brutal. The video-gen know-how was already in our ReAct agent, but nobody could find it. Here is how a UI split unlocked it. weekly 0.6 https://kotonia.ai/zh/articles/features-existed-ux-was-the-bottleneck/ 2026-05-28T14:38:49.696120Z https://kotonia.ai/articles/features-existed-ux-was-the-bottleneck/hero.png 个人开发中"功能齐全却没人用"的问题，靠 UI 拆分解决了 视频生成是新用户主要注册入口但首日流失严重。ReAct agent 中本已包含视频生成工作流但无人触达。本文记录如何用 UI 拆分把已有能力释放出来。 weekly 0.6 https://kotonia.ai/articles/hidream-o1-lora-why/ 2026-05-28T15:21:23.490368Z https://kotonia.ai/articles/hidream-o1-lora-why/showcase2500_04_off_vs_on.png HiDream-O1-ImageのLoRAを自作した理由——包み隠さない版 HiDream-O1-Image向け汎用アニメ/セミリアル質感改善LoRAを作った動機、191枚の手動データ収集、NSFWコンバージョン戦略、そしてnoindex記事がユニーク閲覧1位だった話。技術詳細の日本語版。 weekly 0.7 https://kotonia.ai/en/articles/hidream-o1-lora-why/ 2026-05-28T15:21:23.867836Z https://kotonia.ai/articles/hidream-o1-lora-why/showcase2500_04_off_vs_on.png Why I Trained a LoRA for HiDream-O1-Image — The Unfiltered Story The unfiltered story behind the HiDream-O1-Image LoRA: why anime quality is O1's weak spot, how 191 hand-picked images trained a visual booster, NSFW controllability, and what my analytics revealed about conversion drivers. weekly 0.6 https://kotonia.ai/articles/zenn-canonical-crosspost-seo/ 2026-05-28T15:21:23.526580Z https://kotonia.ai/articles/zenn-canonical-crosspost-seo/cover.png Zennにcanonicalを書いたつもりだったが、実際には効いていなかった話 個人開発サービスの登録数が落ちたので流入とcanonicalを調べ直したら、Zennへの全文クロスポスト運用に穴が見つかった記録。 weekly 0.7 https://kotonia.ai/en/articles/residual-video-domain-i2v/ 2026-05-28T15:21:23.902280Z https://kotonia.ai/articles/residual-video-domain-i2v/cover.png Betting on the video niche the big labs walked away from — model A/B to making I2V the mainstay Is there niche demand in free creation with the guardrails off? A solo dev on one local GPU: from model A/B to making high-res I2V the mainstay. weekly 0.6 https://kotonia.ai/articles/residual-video-domain-i2v/ 2026-05-28T15:21:23.563306Z https://kotonia.ai/articles/residual-video-domain-i2v/cover.png 「大手が降りた動画領域」を個人で実験した記録 — モデル比較から I2V 主力化まで セーフガードを外した自由な創作にニッチ需要はあるのか。個人開発でローカル GPU 一枚、モデル A/B から高解像度 I2V を主力に据えるまでの試行錯誤。 weekly 0.7 https://kotonia.ai/en/articles/hidream-prompt-enhancer/ 2026-05-28T15:21:23.940902Z https://kotonia.ai/articles/hidream-prompt-enhancer/enhanced-anime.jpg HiDream Raw Output Failed → Tried Dev-2604 → VRAM Math Killed It → Won with a Prompt Enhancer Instead How we discovered that HiDream-O1's raw outputs collapse on plain Japanese prompts, tried switching to Dev-2604, found the VRAM math impossible, and won by adding a Gemini Flash Lite prompt enhancer instead. Four non-obvious HiDream pitfalls documented via A/B benchmarking. weekly 0.6 https://kotonia.ai/articles/hidream-prompt-enhancer/ 2026-05-28T15:21:23.601319Z https://kotonia.ai/articles/hidream-prompt-enhancer/enhanced-anime.jpg HiDream のぽん出し → Dev に逃げる → VRAM が破綻 → プロンプトで殴って勝った話 HiDream-O1-Image のぽん出しが日本語プロンプトで崩壊する問題を、Dev-2604 への乗り換え検討 → VRAM 制約で破綻 → プロンプトエンハンサーで解決した実装記録。HiDream 特有の罠 4 つ (ブランド名焼き込み / cute で幼児体型 / Wong Kar-wai で韓国字幕 / idol-class で自動キャプション) を A/B bench で発見。 weekly 0.7 https://kotonia.ai/en/articles/consent-dilemma-10beat-pipeline/ 2026-05-28T15:21:23.982109Z https://kotonia.ai/articles/consent-dilemma-10beat-pipeline/cover.png Turning a 1-Line Idea Into a 40-Second Short with a 10-Beat Local Video Pipeline Full pipeline: Gemma 4 31B expands a one-liner into a 10-beat script, HiDream generates images, LTX-2 I2V renders clips, and ffmpeg assembles everything — all on one local GPU in 25–30 min. weekly 0.6 https://kotonia.ai/en/articles/mesugaki-english-persona-gemini-audio/ 2026-05-28T15:21:24.020163Z https://kotonia.ai/articles/mesugaki-english-persona-gemini-audio/cover.png Building a Sarcastic AI English Tutor with Persona-as-Code and Gemini Audio Input for Pronunciation Correction How I built a high-attitude AI English tutor using persona-as-code design, Qwen3-ASR for multilingual STT, and Gemini audio input for real pronunciation feedback — solo dev perspective. weekly 0.6 https://kotonia.ai/en/articles/ltx2-cold-start-vram-coexistence/ 2026-05-28T15:21:24.055552Z https://kotonia.ai/articles/ltx2-cold-start-vram-coexistence/cover.png Running LTX-2.3 Alongside TTS on a Single 96GB GPU with a Cold-Start Architecture How to go from 86 GiB idle VRAM (instant OOM) to 0 GiB idle / 40 GiB peak by using a cold-start design for LTX-2.3 on one RTX Pro 6000 Blackwell. weekly 0.6 https://kotonia.ai/en/articles/ltx2-22b-fp8-cast-quantization/ 2026-05-28T15:21:24.084364Z https://kotonia.ai/articles/ltx2-22b-fp8-cast-quantization/cover.png Cutting LTX-2 22B Peak VRAM by 40% with fp8_cast — and Why optimum-quanto Was a Trap How fp8_cast reduced LTX-2 22B peak VRAM from 40 GiB to 24 GiB in cold-start mode, and why optimum-quanto silently breaks the transformer. weekly 0.6 https://kotonia.ai/en/articles/hidream-skeleton-pose-prompt/ 2026-05-28T15:21:24.115315Z https://kotonia.ai/articles/hidream-skeleton-pose-prompt/A_baseline.jpg HiDream Skeleton Mode: Prompt Beats OpenPose Ref — 8 Patterns Benchmarked Benchmarking HiDream-O1-Image skeleton mode across 8 patterns reveals 3 counterintuitive findings about openpose refs, resolution drops, and shift values. weekly 0.6 https://kotonia.ai/en/articles/comedy-shorts-claude-gemini/ 2026-05-28T15:21:24.145463Z https://kotonia.ai/articles/comedy-shorts-claude-gemini/cover.png Replicating a Language-Learning Comedy Short with Claude Code — Gemini as a Multimodal Sub-Agent Building a local GPU + Gemini 3.1 Pro hybrid pipeline that generates publishable comedy Shorts from a single line of text in under 60 seconds. weekly 0.6 https://kotonia.ai/en/articles/96gb-vram-mma-and-5-years/ 2026-05-28T15:21:24.186144Z https://kotonia.ai/articles/96gb-vram-mma-and-5-years/case_B_storyboard_local_llm.png Five Years Later, I Finally Have 96GB VRAM — What It Actually Unlocks for Agent Loops Not a GPU unboxing. A real look at what 96GB VRAM enables for multi-model agent pipelines — and where it still hits its limits. weekly 0.6 https://kotonia.ai/en/articles/hidream-quality-speed-bench/ 2026-05-28T15:21:24.214876Z https://kotonia.ai/articles/hidream-quality-speed-bench/presets.jpg HiDream-O1-Image 3–8x Faster: Benchmarking Steps, CFG, and Resolution Real-world timing benchmarks for HiDream-O1-Image Full — tuning steps, guidance scale, and resolution to speed up iteration without killing quality. weekly 0.6 https://kotonia.ai/articles/mesugaki-english-persona-gemini-audio/ 2026-05-28T15:21:23.638783Z https://kotonia.ai/articles/mesugaki-english-persona-gemini-audio/cover.png メスガキ英会話で検索上位を取って、そして失った話 個人開発の niche AI 英会話「メスガキ英語学習」がなぜ刺さったか、検索1・2位を取った戦略、そして自前の原本がZennに奪われた顛末。 weekly 0.7 https://kotonia.ai/articles/consent-dilemma-10beat-pipeline/ 2026-05-28T15:21:23.676035Z https://kotonia.ai/articles/consent-dilemma-10beat-pipeline/cover.png 1 行アイデアを 40 秒コメディ動画にする 10-beat パイプライン ローカル GPU で AI 動画コメディを量産したい開発者向け。1 行のアイデアから LTX-2 + HiDream + Gemma 4 で 40 秒の縦長動画 (Hook + 字幕付き) を 25 分で生成する 10-beat パイプラインの設計、罠、コード構造を記録。 weekly 0.7 https://kotonia.ai/articles/96gb-vram-mma-and-5-years/ 2026-05-26T15:26:29.114514Z https://kotonia.ai/articles/96gb-vram-mma-and-5-years/case_B_storyboard_local_llm.png 3万円のChromebookから、96GBのGPUへ ― 5年越しに諦めなかった話 AIをやりたくても非力なマシンしかなかった5年。Colabでお茶を濁し、震える指でRTX PRO 6000 Blackwell（96GB）を買うまでの、個人開発者の記録。 weekly 0.7 https://kotonia.ai/articles/hidream-quality-speed-bench/ 2026-05-15T23:48:42.299332Z https://kotonia.ai/articles/hidream-quality-speed-bench/presets.jpg HiDream-O1-Image を 3〜8 倍速く使う: steps / CFG / 解像度の実測ベンチ HiDream-O1-Image Full の steps / CFG / 解像度を同一 prompt / seed で振り、生成速度と品質の落とし所を実測した記録。 weekly 0.7 https://kotonia.ai/articles/hidream-skeleton-pose-prompt/ 2026-05-28T15:21:23.713830Z https://kotonia.ai/articles/hidream-skeleton-pose-prompt/A_baseline.jpg HiDream skeleton: openpose ref より prompt が強い (実証 8 パターン) HiDream-O1-Image (8B Full) の skeleton モードを 8 パターン + layout 3 パターンで実証ベンチ。openpose ref を渡すとかえってポーズが固着し、prompt 経由の指定が強い理由を pipeline.py を読み解いて解説。 weekly 0.7 https://kotonia.ai/articles/comedy-shorts-claude-gemini/ 2026-05-28T15:21:23.752473Z https://kotonia.ai/articles/comedy-shorts-claude-gemini/cover.png 言語学習ショート動画を Claude Code で再現してみた — Gemini を sub-agent 化したマルチモーダル拡張 Pingo 風言語学習ショート動画コメディを Claude Code 開発環境でローカル GPU + Gemini 3.1 Pro Preview のハイブリッド構成で再現した話。サブエージェント化でメイン agent の context を膨らませず editorial signal だけフロンティアに委ねる。 weekly 0.7 https://kotonia.ai/articles/ltx2-22b-fp8-cast-quantization/ 2026-05-28T15:21:23.797749Z https://kotonia.ai/articles/ltx2-22b-fp8-cast-quantization/cover.png LTX-2 22B を fp8_cast で peak VRAM 40% 削減した話 — optimum-quanto は罠だった LTX-2.3 22B の量子化を試した記録。optimum-quanto は LTX-2 transformer と互換性問題で動かず、LTX-2 native の `QuantizationPolicy.fp8_cast()` に切り替えて peak VRAM を 40 GiB → 24 GiB（cold-start, 768×512）に圧縮。3 解像度のベンチマークと cold-start / persistent の使い分け判断まで。 weekly 0.7 https://kotonia.ai/articles/ltx2-cold-start-vram-coexistence/ 2026-05-28T15:21:23.836580Z https://kotonia.ai/articles/ltx2-cold-start-vram-coexistence/cover.png LTX-2.3 を 96GB GPU 1 枚で TTS と同居させる cold-start 構成 音声ロールプレイ製品に LTX-2.3 を組み込む際、persistent モードでは 86 GiB を VRAM に置きっぱなしになり TTS/Ditto と同居できない。cold-start に切り替えて idle 0 GiB / peak 40 GiB を達成した実装記録。 weekly 0.7