| v4.6.4 | ## What's New / 更新内容 - Fixed a dialogue extraction bug where character names containing specific particles were incorrectly filtered out. / セリフ抽出時の助詞(「と」「の」等)を含む特定文字長キャラクターのセリフ欠落バグを修正。 - Improved exact cast matching logic. / 完全一致判定で、名前の後に続く発言内容が正しく描画されるように改善。 | High | 6/3/2026 |
| v4.4.3 | ## What's New / 更新内容 ### v4.4.3 (2026-05-28) - **[Fix & Model Sanitization]** OpenAI画像生成APIクライアント(`openai.js`)内に残存していた「gpt-image-2」という架空モデルの定義を完全に排除し、実在モデルである「dall-e-3」に正常化。また不要なデバッグ用モデルリストスクリプト等を削除。 - **[Fix & Model Sanitization]** Completely eradicated fictitious model ID "gpt-image-2" inside openai.js and standardized on "dall-e-3". Cleaned up obsolete diagnostic/test scripts. | High | 5/28/2026 |
| v4.0.3 | ## What's New / 更新内容 ### v4.0.3 (2026-05-21) - ** [Fix & UX] ** 吹き出しへの五感ラベル混入バグおよびアクターZONE・EMOTIONスタイル誤切替の修正 / Fix sensory label hallucination in speech bubbles and incorrect actor ZONE/EMOTION style transitions | High | 5/21/2026 |
| v3.54-alpha | ## What's New / 更新内容 - Completed 360° background integration pipeline. Injected 4 cropped perspective images as multimodal references in Gemini mode for highly accurate per-panel scenery generation. / 360°背景統合パイプライン(Phase 3)の完了と安定化。Geminiモードで4枚のクロップ画像をマルチモーダル参照として注入し、コマ別の方角に合わせた精密な背景描画を実現。 - Fixed UI layout issue where text and image overlays overlapped in the camera work preview panel. / カメラワーク確認UIにおいて、テキスト(「各コマで使用される背景の方角」)と画像オーバーレイが重なって表示されるレイアウト崩れを修正。 | High | 5/15/2026 |
| v3.12-alpha | ### v3.12-alpha (2026-05-09) - **[UI/UX]** UI/ログの完全日本語化を実施。思考トレース(Thinking Trace)が利用できない標準モデル使用時のメッセージを「通常処理が完了しました。」と簡潔化し、不要な注釈を排除しました。 / Localized UI logs to Japanese completely. Simplified the fallback message for standard models without Thinking Trace to "通常処理が完了しました。" (Normal processing completed), removing confusing annotations. | High | 5/9/2026 |
| v3.02.0-alpha | ## What's New / 変更点 ### BugFix - **[BugFix]** Fixed a bug in ChatGPT mode (GPT-image-2) where a black border was drawn around the entire page. The phrase "OUTER FRAME" in the CANVAS LOCK section was interpreted by DALL-E as a physical border drawing instruction; replaced with "image boundaries" and added explicit "NO PAGE BORDERS" rule to the ANTI-NOISE section as a double-block countermeasure. - **[BugFix]** ChatGPTモード(GPT-image-2)で生成されるプロンプトにおいて、ページ全体に黒い外枠(border)が描画される問題を修正。CANVAS LOCKセクションの | High | 5/3/2026 |
| v2.77.0-alpha | # v2.77.0-alpha — Enhanced GPT-image-2 Anti-Noise Protocol / GPT-image-2ノイズ対策プロトコル大幅強化 ## What's New / 変更点 ### ✨ Feature / 新機能 - **Enhanced GPT-image-2 Anti-Noise Protocol**: Integrated comprehensive noise countermeasures based on community research to significantly improve image quality when using ChatGPT Images 2.0 mode. - **Positive Safe Rendering Block**: Injected safe rendering keywords (`smooth cel shading`, `clean color surfaces`, `low texture density`, etc.) to actively guide the mo | High | 4/29/2026 |
| v2.66.0-alpha | ### 📝 Update Documentation for ChatGPT & Enhancements #### English * **[Docs]** Updated `README.md` to formally document the new **ChatGPT Images 2.0 Mode** and **Scenario Enhancement** features. * Adjusted terminology to reflect compatibility with DALL-E 3, enforcing A4 aspect ratios and vertical text rules via prompt injection. #### 日本語 * **[Docs]** `README.md` を更新し、新たに搭載された **ChatGPT画像生成サポート** および **シナリオ演出強化** 機能に関する公式ドキュメントを追加しました。 * DALL-E 3生成に最適化されたA4縦長指定や、縦書きの視線誘導フォーマットに対応した設計であることを技術詳 | High | 4/22/2026 |
| v2.60.0-alpha | # Nano Banana Pro v2.60.0-alpha 🚀 **Dynamic Camera Recovery & Scenario Normalization Update** ## 🔧 Fixes & Optimizations * **Camera Lens Mapping [Critical Fix]**: Fixed an issue where Extreme Camera Angles (v2.55) were being stripped of their specific lens distortion weights (fish-eye, telephoto, dutch angle, etc.) during generation. AI-selected camera names are now explicitly mapped to heavy weighted lens tags, restoring the hyper-dynamic camera distortion. * **Scenario Enhancements Nor | High | 4/21/2026 |
| v2.59.0-alpha | ## Nano Banana Pro v2.59.0-alpha ### EN - **Optimization**: Completely overhauled the Gemini API model fallback lists in preparation for the deprecation of the official Imagen series models (June 24, 2026). Promoted `gemini-3.1-flash-lite-preview` and `gemini-2.5-flash-image` as primary free-tier fallbacks, and correctly updated modal payloads for native Gemini image generation. - **System**: Implemented the multi-agent `Model Audit Protocol` across the standard `AGENTS.md` file, standardizing | High | 4/21/2026 |
| v2.58.0-alpha | ## Nano Banana Pro v2.58.0-alpha ### EN - **Fix**: Resolved issue where low-angle (Aori) and worm's-eye camera shots sometimes generated shoe-only close-up panels instead of full-body character compositions. All low-angle prompts now explicitly mandate full-body rendering. - **Fix**: Removed "cracking/shattering panel borders" VFX from the IMPACT emotion style, which caused undesirable fracture lines on panel edges. Replaced with a clean "intense glowing energy aura" effect. ### JP - **修正**: ロ | High | 4/20/2026 |
| v2.57.0-alpha | ## Nano Banana Pro v2.57.0-alpha ### EN - **Enhancement**: Refined Fisheye (Ultra wide-angle) camera perspective prompt logic. - Added strict prohibition against generating "bulbous noses" (pig noses) during fisheye distortion to prevent undesirable facial collapse while maintaining extreme visual perspective. ### JP - **改善**: フィッシュアイ(超広角)アングル時のカメラプロンプト指示を修正しました。 - 極限の遠近感を出しつつも、「団子鼻(豚鼻)」になるような不細工な造形崩壊をシステム側で明示的に絶対禁止とするよう対応しました。 | High | 4/20/2026 |
| v2.56.0-alpha | ### Changes / 更新内容 - **[Fix]** Removed "Macro Close-up" from camera options to prevent the AI from generating huge face/eye close-ups that violate the "all characters must be fully drawn" rule. - **[Feature]** Replaced the macro option with "Panning/Tracking Shot" where the camera tracks characters with dynamic motion blur. --- - **[修正]** カメラ選択肢から「マクロ特写」(顔や瞳の極端なアップ)を削除しました。これにより、AIが全員を描画せずに画面全体に巨大な顔を生成してしまうグリッチを防止します。 - **[機能追加]** マクロ特写の代わりに、「パンニング/追跡ショット」(キャラを追跡し背景がモーションブラーで流れる演出)を追加しました。 | High | 4/20/2026 |
| v2.55.0-alpha | ## What's New / 変更内容 ### 🎬 Extreme Physical Camera Description Mode (v2.55) - **Initial scenario generation now produces 3x-enhanced-level camera descriptions automatically** — no need to press the Enhance button multiple times - Camera Direction Tag section in scenario prompt completely rewritten: AI must now describe **physical scene details** (what the camera sees, how objects distort, how light falls) instead of just naming the camera angle - Added **Extreme Reaction Mandate**: all charact | High | 4/20/2026 |
| v2.54.0-alpha | # Hyper-Dynamic Camera Protocol Maximize - Boosted AI image generation prompt weights for camera distortion and angle enforcement to their absolute maximum limits (2.5 - 3.0 range). - Re-implemented the `getCameraForPanel` AI angle classification logic to correctly force specific camera distortions based on AI's cinematic choice. - Added a Japanese watermark on the outside bottom-left border of the 4th panel for automated tracking. # 超ダイナミックカメラ・プロトコル最大化 - カメラの歪みとアングル強制のAI画像生成プロンプトのウェイトを、限界突破の絶対 | High | 4/20/2026 |
| v2.53.0-alpha | ## What's New / 変更点 ### 🎬 ANTIGRAVITY HYPER-DYNAMIC CAMERA PROTOCOL v3.2 ### 🎬 カメラワーク超強化プロトコル v3.2 **EN:** - Integrated the Antigravity Hyper-Dynamic Camera Protocol v3.2 — a comprehensive overhaul of manga panel composition to enforce extreme, non-eye-level camera angles in every panel. - Added **OVERRIDE PROHIBITION** to prevent the AI from softening distortion for "readability" or "narrative flow." - Added **MANDATORY LENS ENFORCEMENT** blocks to each panel requiring visible proof of dist | High | 4/19/2026 |
| v2.52.0-alpha | ### 🇯🇵 日本語 - **[BugFix]** PRO TIPSの文章における「カンニングできる」という不適切な表現を「参照できる」に修正 ### 🇺🇸 English - **[BugFix]** Corrected inappropriate wording ('cheat') in PRO TIPS description to 'reference'. | High | 4/18/2026 |
| v2.51.0-alpha | ## 🎯 Scenario-Driven Location & Outfit (v2.43 Architecture) ### English - **Location & Outfit now read from scenario text** — no more STEP 2 re-generation needed to change location/outfit - Input fields serve as convenience shortcuts for STEP 2; afterwards, edit `Location:` / `Outfit:` lines directly in the scenario - GENERATION PREVIEW now shows real-time values parsed from scenario text - Outfit placeholder updated with "普段着" (casual wear) example - All existing outfit enforcement (per-panel | High | 4/16/2026 |
| v2.50.0-alpha | ## 🇯🇵 日本語 - **不具合修正 (Bug Fix)**: 新しいシナリオを生成した際に、前回のシナリオ強化状態(元のシナリオ、強化ログ)がリセットされず、「強化済」の表示が残ったままになったり、「強化前に戻す」を実行した際に古いシナリオが復元されてしまう問題を修正しました。シナリオ生成開始時に確実に状態をクリアするようになりました。 ## 🇺🇸 English - **Bug Fix**: Fixed an issue where the Scenario Enhancement state (original scenario, enhancement logs) was not properly reset when generating a new scenario. This caused the UI to incorrectly display the "Scenario Enhanced" status and the "Revert" button to unexpectedly restore a scenario from a previous s | High | 4/15/2026 |
| v2.49.0-alpha | # v2.49 Alpha - API認証バッジ & 描画エリアロック ## 🇯🇵 日本語 ### 新機能・改善 - **API認証状態バッジ**: タイトル枠内に「✅ API認証済」/「⚠ 未接続」バッジを常時表示 - 緑色光るドット付きバッジ(認証済み時) - 赤色点滅バッジ(未接続時) - **サブタイトル中央配置**: 「SOCIAL SATIRE ENGINE」とAPI認証バッジを水平中央揃えに修正 - **描画エリアロックオーバーレイ**: STEP3(プロンプト構築)未完了時、STEP4の画像生成結果エリア全体にぼかしオーバーレイを表示 - 枠線・タイトルヘッダー・中央アイコン・「Ready to Start」テキストがボケた状態に - STEP3完了後に初めてクリアに表示される --- ## 🇺🇸 English ### New Features & Improvements - **API Authentication Badge**: Persistent "✅ API Authenticated" / "⚠ Not Connect | High | 4/15/2026 |
| v2.47.0-alpha | # v2.47 Alpha - Camera Work & Dialogue Enhancement / カメラワーク・セリフ強化 ## 🐛 Bug Fixes / バグ修正 - **STEP3 "最終プロンプトを構築する" ボタン押下時にSTEP2エリアが消えるバグを修正** - Fixed: STEP2 overlay incorrectly activated during STEP3 prompt assembly (`isAssembling` removed from STEP2 lock condition) - **STEP3構築中にSTEP4エリア(ボタン下)が真っ黒になりログが見えないバグを修正** - Fixed: STEP4 overlay now hides during prompt assembly so users can see the build progress log ## ✨ New Features / 新機能 ### シナリオ強化パネル - 2カテゴリ追加 (4→6) - **📷 カメラワーク強化**: 各コマにアオリ | High | 4/15/2026 |
| v2.46.0-alpha | ## Nano Banana Pro v2.46 Alpha - UI Locking Mechanism Overhaul 🔒 ### 🐞 Bug Fixes & Improvements (English) * **STEP UI Disablement Overhaul**: Completely rewrote the UI element deactivation logic for steps that are currently locked (e.g. while processing or waiting for preceding steps). * Replaced unreliable Tailwind `blur` filter classes with robust inline-styled overlay panels (`absolute inset-[-2px] z-[200] background-color: rgba(0,0,0,0.92) backdrop-filter: blur(8px)`). * Th | High | 4/14/2026 |
| v2.45.0-alpha | ## Release Notes (v2.45 Alpha) ### 🎨 UI & UX Improvements (UI/UXの改善) - **Standardized Progress Windows**: Unified font sizes across all progress logs and explanation panels (STEP 4 & Content Policy) to `text-xs` (12px) for improved readability, while maintaining `11px` for high-density neural process logs. Fixed a bug where dynamic height changes caused UI layout shifts by enforcing a static `160px` height structure. (進捗窓のフォントサイズと高さを全ステップで統一。コンテンツポリシーやPRO TIPの説明は視認性の高い12pxに、プロセスログは情報密度の高い11 | Medium | 4/14/2026 |
| v2.44.0-alpha | ## 🚀 Updates / 変更点 * **UI/UX Refinements (Progress Trackers):** Fixed intermittent countdown stalling during cross-model fallback steps by switching interval logic to a resilient inline regex update. Also decoupled `isGeneratingImage` state in STEP 4 to prevent total screen deactivation while retaining action-lock on necessary elements. * **Content Policy Rescue UI Upgrade:** Frozen the generated output console via inline `style={{ height: '240px' }}` forcing stable panel footprint regardless | Medium | 4/14/2026 |
| v2.43.0-alpha | ## 🚀 Changes in v2.43.0-alpha / 変更点 ### 🛡️ Anti-Cloning Logic Enhancement / キャラクター重複防止の強化 - **Total Character Count Protocol (登場人数の厳密な制約)**: - The AI generation prompt now explicitly calculates and locks the total number of individuals in the scene. - If a scene has 2 main actors and 1 background character, it will strictly enforce "EXACTLY 3 distinct individuals. Each person appears ONLY ONCE." This forcefully stops the AI from mistakenly duplicating main characters as background observ | High | 4/14/2026 |
| v2.41.0-alpha | ### Changes / 最新の変更点 * **New Feature: Scenario Enhancement Panel** — Added a collapsible panel between STEP 2 output and STEP 3 that allows fine-tuning the generated scenario's dramatic impact before image generation. * 4 individual toggle switches: **Expressions** (大げさ表情), **Body Language** (全身リアクション), **Effects** (照明・VFX演出), **Backgrounds** (背景描写強化) * "Execute Enhancement" button sends the scenario back to Gemini for dramatic upgrade * "Undo" button restores the original pre-enhancement | Medium | 4/13/2026 |
| v2.40.0-alpha | ### Changes / 最新の変更点 * **Revert (Camera Experiments v2.36–v2.39)**: Rolled back 4 consecutive versions of camera angle experiments that failed to break the eye-level prior of Gemini's image generation model, while degrading overall prompt quality. Restored the stable v2.35 prompt structure. * **Restore (Panel Grid)**: Re-enabled `CRITICAL PANEL SIZE COMMAND` (strict 4-equal-panel grid) and `CRITICAL COMPOSITION RATIO` (2:3 golden ratio) that were removed during camera experiments. * **Restore ( | Medium | 4/13/2026 |
| v2.39.0-alpha | ### Changes / 最新の変更点 * **Fix (Markdown Prior Override)**: Imagen often ignores structural Markdown tags like `[SCENE CAMERA OVERRIDE]:` as pseudo-code or structural metadata, defaulting back to flat comic-strip layouts. The camera angle instruction has been fully integrated into the natural language of the `Visual Action` description as a strong declarative sentence. * **Refactor (Relax 4-Koma Grid)**: Removed the `CRITICAL COMPOSITION RATIO` (Strict 2:3 golden ratio) entirely and relaxed the ` | Medium | 4/13/2026 |
| v2.38.0-alpha | ### Changes / 最新の変更点 * **Fix (Camera Angle Deadlock)**: Imagen 4 was stubbornly defaulting to a flat eye-level camera because the `CHARACTER BODY POSITION LOCK` rules logically forced a theatrical "stage-play" layout (e.g., "Physically standing on the LEFT half"). * **Refactor (Spatial Relaxation)**: Changed the position lock rules from absolute physical stage instructions (`MUST be physically standing...`) to camera-relative visual composition instructions (`From the camera's perspective... a | Medium | 4/13/2026 |
| v2.37.0-alpha | ### Changes / 最新の変更点 * **Fix (Camera Override)**: Imagen 4 was aggressively ignoring camera angles due to the strict "4 EQUAL horizontal panels" and "golden ratio" alignment instructions, rendering flat eye-level grids. Modified the Composition Ratio rule to explicitly permit "warping the interior perspective". * **Refactor (Camera Prompt)**: Converted the bracketed `[Camera Angle: ...]` tag entirely into natural language (`The scene is shot from a ...`). Removed the flat "Medium-Full Shot" and | Medium | 4/13/2026 |
| v2.36.0-alpha | ### Changes / 最新の変更点 * **Fix (Prompt)**: Resolved an issue where Imagen 4 ignored the `Camera` instruction due to pseudo-code formatting. By embedding the dynamic camera angle directly into the `Visual Action` text block, the AI now successfully renders dramatic shots (Low Angle, Bird's Eye, Dutch Angle) instead of defaulting to monotonous eye-level perspectives. * **Refactor**: Cleaned up orphaned code variables (`VAR_PANEL_1_KI`, etc.) that were no longer used in the layout assembly. --- * | Medium | 4/13/2026 |
| v2.35.0-alpha | ## 🚀 Changes in v2.35.0-alpha /変更点 ### 🛡️ Content Policy Rescue / コンテンツポリシー救済パネルの強化 - **Smart Collapsible UI (直感的な折りたたみUI)**: - The Rescue Panel now automatically expands only when an AI Content Policy Error is detected, maintaining a clean UI without unnecessary whitespace. - (画像生成が安全基準でブロックされた時だけパネルが自動展開される、完全折りたたみUIを実装。不要な余白をゼロにしました) - **Prompt Integrity Lock (修正プロンプトの完全維持)**: - The AI prompt advisor is now strictly guided via Meta-Prompting to preserve the exact length, detail, and | Medium | 4/12/2026 |
| v2.34.0-alpha | ## 🚀 Changes in v2.34.0-alpha /変更点 ### 🎭 Manga Composition Variety / 登場人数のバラエティ対応 - **Dynamic Character Counts (コマごとの話者数の意図的な変化)**: - Enforced a rule to naturally mix the number of speaking characters across the 4 panels. At least one panel MUST be a solo monologue, and at least one MUST be a 2-person dialogue, preventing the AI from creating 4 identical panel compositions. - (全コマが同じ人数にならないよう、1人での独白シーンと2人での会話シーンを最低1つずつ入れるルールを義務化しました) - **Background Mob Detection (非話者キャストの背景モブ描画)**: - Ch | High | 4/9/2026 |
| v2.33.0-alpha | ## Changes in v2.33.0-alpha / 変更点 ### Dialogue Punctuation System Overhaul / セリフ句読点システム全面改修 - **Natural sentence-ending expressions**: Replaced the rigid period-only rule with a comprehensive punctuation system that matches character emotions and personality - **Terminal vs Modifier marks**: Classified punctuation into terminal marks (。!?♪♡っ!?) that can end sentences alone, and modifier marks (……~ー) that require a terminal mark after them - **Anti-repetition rule**: Same ending expression canno | High | 4/8/2026 |
| v2.32.0-alpha | ## What's Changed / 変更内容 ### Bug Fixes / バグ修正 - **Fix excessive commas in speech bubbles / 吹き出しの読点過剰挿入を修正** - Removed forced comma insertion rule from both scenario generation (Gemini) and image generation (Imagen) prompts - Manga speech bubbles now use natural Japanese punctuation (periods only, minimal commas) - シナリオ生成・画像生成の両方のプロンプトから「読点を必ず入れよ」ルールを削除 - 漫画の吹き出しに自然な句読点(句点のみ、読点は最小限)を使用するよう変更 - **Fix style labels rendered as text on panels / スタイルラベルがコマ内にテキスト描画されるバグを修正** - Changed 'ART | Medium | 4/6/2026 |
| v2.31.0-alpha | ## What's Changed / 変更内容 ### 🔧 Bug Fixes / バグ修正 **Character Identity Preservation / キャラクター一貫性の強化** - 🔒 BLANK (白目) style now preserves glasses on glasses-wearing characters / BLANK(白目)スタイルで眼鏡キャラの眼鏡が消えなくなりました - 🔒 CHIBI_GAG style now preserves glasses in super-deformed form / ちびデフォルメでも眼鏡が保持されるようになりました - 💇 Hairstyle extraction now prioritizes distinctive styles (twintails, hime cut) over generic (long hair) / 髪型抽出がツインテール等の特徴的スタイルを優先するようになりました - ✨ GLITTER style no longer causes hair length chan | Medium | 4/6/2026 |
| v2.30.0-alpha | ## 🚀 Changes in v2.30.0-alpha /変更点 ### 🎨 Character Consistency Logic / キャラクター同一性の強化 - **Anti-Cloning Enforcement (クローン防止プロトコルの強化)**: - Characters specified in solo-shots will NEVER duplicate. Explicit SINGLE CHARACTER rules are enforced to prevent AI hallucination. - (1人だけのコマでAIが勝手に話し相手を捏造・増殖させる問題を防ぐため、「SOLO SHOT」ルールを追加しました) - **Gag & Emotion Text Filtering (オノマトペやギャグSFXの吹き出し化防止)**: - Visual actions ending in Japanese SFX (e.g., ズコー, ドッ, ガーン) are now cleaned to prevent them from being d | Medium | 4/6/2026 |
| v2.29.0-alpha | ## What's Changed / 変更内容 ### Bug Fixes / バグ修正 - **Glasses Detection (Identity Matrix)**: Rewrote glasses detection logic to use WEIGHTS tags as the authoritative source. Added a locking mechanism to prevent subsequent lines from overwriting the correct determination. Characters without glasses are no longer incorrectly assigned 'MUST HAVE glasses'. - **眼鏡判定ロジック根本修正**: WEIGHTSタグを権威ソースとする方式に全面書き換え。ロック機構により後続行の上書きを防止。眼鏡なしキャラに「MUST HAVE glasses」が誤付与される問題を解消。 - **Narration Leak Prevention (Primary | Medium | 4/5/2026 |
| v2.28.0-alpha | ## v2.28 Alpha - Prompt Quality Overhaul / プロンプト品質大改修 ### What's Changed / 変更点 **🔴 Critical Fixes / 重大バグ修正** - **Glasses Detection Logic Fix**: The Identity Matrix no longer misidentifies characters with glasses (e.g. `度なし` matching `なし`) — Rin and Hikari now correctly get `MUST HAVE glasses` - **眼鏡判定ロジック修正**: Identity Matrixが「度なし」等の表記で眼鏡キャラを「眼鏡なし」と誤判定するバグを修正。リンとヒカリの眼鏡が正しく描画されるように - **EMOTION Tag Speaker Contamination Fix**: `[EMOTION: RETRO]` tags were being extracted as character names in | Medium | 4/4/2026 |
| v2.27-alpha | ## English - Enhanced Safety Age-Up Filter: Automatically replaces sensitive character attributes (girl, teenager, school uniform, serafuku, etc.) with safe adult alternatives in the prompt pipeline. - STEP1 Template Root Fix: Analysis prompt now outputs safe tags (woman, young adult, formal outfit) by default, reducing reliance on post-processing filter. - Character Body Position Lock: Uses visual trait anchoring (hair color, features) to lock character physical positions (LEFT/RIGHT) per pa | Medium | 4/3/2026 |
| v2.25.0-alpha | ## What's New / 変更内容 ### Identity Matrix & Emotion Style Engine / アイデンティティ・マトリクス&感情スタイルエンジン - Character identity locking (hair color, hairstyle, glasses) enforced across all 4 panels / キャラクターの識別特徴(髪色・髪型・メガネ有無)を4コマ全パネルで強制固定 - 12 emotion style tags (`[EMOTION: XXX]`) for dynamic art direction per panel / パネルごとに12種の感情絵柄タグで画風を動的に制御 - NORMAL, CHIBI_GAG, GEKIGA, SHOUJO, HORROR, BLANK, IMPACT, WATERCOLOR, SKETCH, RETRO, GLITTER, SHADOW ### API Model Priority Update / APIモデル優先順位の更新 - Promote | Medium | 4/1/2026 |
| v2.24.0-alpha | ### Changes - Added ANTI-FLOATING-EYE RULE to prevent floating close-up eyes or partial face crops from appearing as ghostly background overlays in panels. - Removed "Dramatic Close-up on eyes" from camera angle pool, replaced with "Medium Shot (Waist-up)". - Added CRITICAL COMPOSITION BAN constraint to Important Constraints section reinforcing the anti-floating-eye rule. - Strengthened horizontal watermark enforcement: explicitly prohibits 90-degree rotation, vertical stacking, and per-lett | Medium | 3/25/2026 |
| v2.23.0-alpha | ## What's Changed / 変更内容 ### Model Stack Modernization / モデルスタック最新化 **Removed deprecated models / 廃止モデルの除去:** - ❌ `gemini-2.0-flash` — Returns 404 (no longer available) - ❌ `gemini-2.0-flash-lite` — Returns 404 (no longer available) **Added next-gen preview fallbacks / 次世代プレビューモデルをフォールバックに追加:** - 🆕 `gemini-3-flash-preview` — Next-gen fast model (preview) - 🆕 `gemini-3.1-flash-lite-preview` — Next-gen lightweight model (preview) ### Updated Model Priority / 更新されたモデル優先順位 1. `g | Medium | 3/23/2026 |
| v2.22.0-alpha | ## What's Changed / 変更内容 ### English - **Fixed**: Critical bug where all speech bubbles were suppressed during prompt generation - Root cause: `extractDialogueOnly()` function failed to match character names when format differed between castList and scenario (e.g. "アカリ (AKARI)" vs "アカリ") - When matching failed, the function returned a destructive fallback: `"(CRITICAL: Do NOT draw any speech bubbles or text)"`, which caused the image generation AI to omit all dialogue - **Improved**: | Medium | 3/23/2026 |
| v2.21.0-alpha | ## What's Changed / 変更内容 ### English - **Fixed**: Local development (localhost) STEP2 scenario generation failing with "SYSTEM FAILURE: All models failed" - **Root Cause**: Google Search Grounding tool calls are blocked/restricted from `localhost` origins, and the failover condition in `gemini.js` only caught `400` and `"not supported"` errors — missing other error types (403, CORS, PERMISSION_DENIED) - **Fix**: Expanded grounding failover to retry WITHOUT tools on ANY error when `googleSe | Low | 3/19/2026 |
| v2.20.0-alpha | ## What's Changed / 変更内容 ### English - Added `CROSS-PANEL OUTFIT CONSISTENCY (MANDATORY)` rule to the image generation prompt - Ensures every character wears the EXACT same outfit across all 4 panels - Works regardless of whether a custom outfit override is specified - Placed immediately after the `Character Identity Anchor` section for maximum effectiveness ### 日本語 - 画像生成プロンプトに `CROSS-PANEL OUTFIT CONSISTENCY (MANDATORY)` ルールを追加 - 全4コマでキャラクターの服装を厳密に一致させるための指示文 - カスタム服装指定(Outfit Ove | Low | 3/18/2026 |
| v2.19.0-alpha | ## What's Changed / 変更内容 ### 🛡️ Outfit Override Bias Fix (Plan 3: Dual Defense) / 服装オーバーライドのバイアス修正(案3: 二重防御) When a custom outfit (e.g. spacesuit) was specified via OUTFIT OVERRIDE, the AI image generator would revert to the reference image's default clothing (e.g. school uniform) in panels 2-4, caused by the reference image's visual bias overpowering the text instruction. / カスタム服装(例: 宇宙服)をOUTFIT OVERRIDEで指定した際、参照画像の視覚バイアスにより2〜4コマ目でデフォルト服装(例: 制服)に戻る問題を修正。 **Fix 1 - REFERENCE IMAGE CLO | Low | 3/14/2026 |
| v2.18.0-alpha | ## What's Changed / 変更内容 ### API Model Optimization / APIモデル最適化 - **Text Generation (gemini.js)**: Replaced deprecated/unstable models with stable GA versions: - `gemini-2.0-pro-exp` → `gemini-2.0-flash` (Stable GA) - `gemini-pro-latest` → `gemini-2.5-flash-lite` (Explicit lightweight) - `gemini-flash-latest` → `gemini-2.0-flash-lite` (Explicit lightweight) - **Image Generation (imagen.js)**: Removed dead Imagen 2 legacy model (`image-generation-001`) - **テキスト生成 (gemini.js)**: 廃 | Low | 3/14/2026 |
| v2.16.0-alpha | ## Release Notes (v2.16.0-alpha) ### 🚀 Features & Improvements (English) * **Relaxed Cast Limits**: Relaxed the strict character exclusion logic in the background, allowing non-speaking characters and background mobs to be generated naturally alongside the main speaking characters. This restores the lively and expressive nature of the manga panels. * **Anti-Hallucination Measures**: Implemented a highly restrictive negative prompt `(ABSOLUTELY NO ENGLISH TEXT outside watermark. Do NOT | Low | 3/7/2026 |
| v2.11_Alpha | ### 🚀 Updates (v2.11 Alpha) - **Watermark Text Fix**: Optimized the watermark generation prompt with strict quote syntax to correctly handle target text without hallucinating internal instructions (e.g., "(LTR) ONLY ONE"). - **Font Constraint Enforcement**: Enhanced speech bubble text rendering stability by explicitly applying the `bold sans-serif Japanese font` constraint to dialogue rendering. ### 🇯🇵 更新内容 (v2.11 Alpha) - **ウォーターマーク文字化け問題の修正**: AIがシステム側の生成指示文を透かしテキストとして誤認("(LTR) ONLY O | Low | 3/7/2026 |
| v2.10.0-alpha | - Fixed the issue where AI generated hallucinations in the watermark text / AIがウォーターマークのテキストでハルシネーションを起こす問題を修正 - Enforced single horizontal rendering for the watermark / ウォーターマークを横書きで1回のみ描画するようルールを強制 - Removed the word "System" from the watermark text to prevent typographic errors / タイポを防ぐためウォーターマークのテキストから"System"の文字を物理的に除外 | Low | 3/6/2026 |
| v2.09.0-alpha | ### Bug Fixes * **Resolved Outfit Hallucination**: Fixed an issue where original character outfit tags (e.g., "school uniform") persisted in the final generation prompt even when a specific outfit override (e.g., "casual wear") was requested. This ensures that the AI accurately reflects the newly specified outfit without mixing clothing styles. ### バグ修正 * **服装のハルシネーション修正**: 「普段着」などの服装の上書きを指定した際、元のキャラクターシートにあった「制服」等の服装指示がプロンプトの末尾にそのまま残存してしまい、AI生成時に両方のデザインが混ざる(ハルシネーションが起きる)問題を修正しました。これにより、指定し | Low | 3/6/2026 |
| v2.08.0-alpha | En (English) Model Tier Optimization: Modernized the API substrate by replacing deprecated models with the latest stable iterations (gemini-2.5-flash, gemini-2.5-pro, gemini-2.0-pro-exp) to enforce zero-failure API connections and resolve 404 Not Found errors. Neural Process Logging: Implemented fully localized (Japanese) "Thinking Process" commentary. Error events (Safety, 404, Output Empty) and transition success messages are now broadcast in real-time to the UI log window. Quality & Safety | Low | 3/6/2026 |
| v2.07-alpha | This update addresses critical API 404 errors causing models to fall back, and fixes an issue where Gemini 2.0 Flash would drop characters when multiple images were uploaded. このアップデートは、APIの404エラーによるモデル転落問題と、複数画像アップロード時にGemini 2.0 Flashがキャラクター解析を途中で打ち切ってしまう重大なバグを修正するものです。 🐛 **Bug Fixes & Improvements / バグ修正と改善** * **[Fix] Character Parsing Drop-off / 複数キャラ解析時の打ち切りバグ修正**: Resolved an issue where Gemini 2.0 Flash would intentionally stop parsing after the first two characters. Altere | Low | 3/4/2026 |
| v2.06-alpha | ## 🚀 Release Notes / リリースノート (v2.06 Alpha) This update addresses critical deployment pathing issues specifically targeted for environments beyond GitHub Pages, such as Hugging Face Spaces. このアップデートは、GitHub Pages以外の環境(Hugging Faceなど)での表示バグを根本的に解決するものです。 ### 🐛 Bug Fixes & Improvements / バグ修正と改善 - **[Fix] Universal Path Resolution / ユニバーサルパス解決**: Changed the Vite build configuration to use a relative base path (`base: './'`). This ensures that all generated assets resolve correctly rega | Low | 2/28/2026 |
| v2.05-alpha | 🚀 Updates / 更新内容 Fixed critical bugs in the automated versioning script that caused infinite loops of "NaN" generation and corrupted version strings across the app. バージョン情報を自動更新するシステムのロジックを根本から書き直し、「NaN(数値ではない)」というエラー文字列が混入するバグや、小数点第3位が勝手に付与されてしまう問題を完全に修正しました。 🔧 Feature Restorations & Fixes / 復旧・修正された機能 * **Strict Version Format Enforcement** * Completely rewrote `update_version.js` to strictly enforce the addition of +0.01 without semantic patches, eliminating floating-point math | Low | 2/28/2026 |
| v2.02-alpha | 🚀 Updates / 更新内容 Fixed layout rules to prevent the 4 manga panels from shrinking and creating excessive white margins. Added a strong anti-cloning protocol to stop the AI from drawing the same character twice in a single panel. 4コマの枠が小さくなりすぎて不要な余白が広がる問題と、AIが同じコマの中に同一キャラクターを2人描いてしまう「分身現象」を抑制する強力なプロンプト対策を実装しました。 🔧 Feature Restorations & Fixes / 復旧・修正された機能 * **Strict Panel Width and Margin Fix / コマ幅と余白の厳格化** * Changed the layout prompt to force panels to occupy 95% of the canvas widt | Low | 2/28/2026 |
| v2.01-alpha | 🚀 Updates / 更新内容 Implemented a strict English prompt protocol ("CRITICAL VISUAL REPRODUCTION PROTOCOL") to forcefully dictate that the AI must maintain 100% fidelity to the attached character reference image, especially concerning hairstyle, hair color, and glasses. 添付画像がある場合、AIに対して「元画像のキャラクターデザイン(髪型、髪色、眼鏡の有無)を100%忠実に再現すること」を強制する、非常に強固な英語のプロトコルを追加実装しました。 🔧 Feature Restorations & Fixes / 復旧・修正された機能 * **Strict adherence to Image Prompting / 画像プロンプトへの追従性を大幅向上** * Placed a stern instr | Low | 2/28/2026 |