GenMedia 1.26.1004

繁體中文

這次更新帶來更多圖片與有聲影片生成選擇，並讓大量素材的操作更流暢，維持原有介面與操作方式。

• Qwen 2.1 Turbo：以 6 步生成圖片或編輯單張圖片，可與原有一般 Profile 切換。
• 提示詞增強：Qwen PE 將簡短描述展開為詳細提示詞，分別支援文生圖與圖生圖，保留原始指令、指定尺寸與 Seed。
• LTX-2.5 有聲影片：新增實驗性 Distilled Q4 文生影，輸出同步立體聲音訊，採 8 步生成及 3 步細化。
• 更多低步數影片選擇：新增 H3 LightX2V Turbo v1.2 的 4 步 LoRA，支援完整與 Pruned FL2VA 模型。
• 更順暢、更省暫存記憶體：改善大量素材的分頁處理、範圍選取與 Profile 篩選，加快音樂 WAV 輸出，減少模型載入與生成時的重複運算。
• 更可靠的工作接續：重新開啟工作區時保留一般／Turbo／提示詞增強 Profile，改善生成完成及資源釋放時的等待問題。

所有新增模型均採本機 Swift／MLX 推論，不需安裝 Python。模型與 LoRA 權重請在模型中心另行下載；提示詞增強會增加前處理時間。

實驗功能範圍：Qwen Turbo 已驗證 256×256 生成與改色編輯；LTX-2.5 已確認 256×256、9 幀影音輸出，建議 32 GB 以上記憶體。H3 v1.2 的完整影片效果，以及 LTX-2.5 的長片、高解析度與音訊品質仍待驗證。低步數不會降低主模型的記憶體需求；已知限制見驗證紀錄。

安裝：下載 GenMedia-1.26.1004-arm64.dmg，開啟後將 GenMedia 拖入 Applications。此 Apple Silicon 安裝包已完成 Developer ID 簽章、Apple 公證與 Gatekeeper 驗證。模型權重不包含在安裝包內。

Qwen 使用指南 · LoRA 使用指南 · 完整更新紀錄

English

This update adds more image and video-with-audio options and makes large asset collections easier to work with, while preserving the existing interface and controls.

• Qwen 2.1 Turbo: Generate or edit a single image in 6 steps, with the standard profiles still available.
• Prompt enhancement: Qwen PE expands short instructions for image generation and editing while preserving your original prompt, selected size, and seed.
• LTX-2.5 video with audio: Experimental Distilled Q4 text-to-video with synchronized stereo audio, using 8 generation steps and 3 refinement steps.
• More low-step video options: H3 LightX2V Turbo v1.2 adds a 4-step LoRA for full and Pruned FL2VA models.
• Smoother workflows and less temporary memory: Faster workspace reconciliation, range selection, profile filtering, and music WAV export, with less repeated work during model loading and generation.
• More reliable session continuity: Workspaces retain standard, Turbo, and prompt-enhanced profile choices when reopened; generation completion and resource cleanup handle process exits more reliably.

All additions run locally with native Swift/MLX; no Python installation is required. Download models and LoRAs separately through Model Center. Prompt enhancement adds preprocessing time.

Experimental scope: Qwen Turbo was checked at 256×256 for generation and color editing. LTX-2.5 video/audio output was checked at 256×256 with 9 frames; 32 GB or more memory is recommended. Full H3 v1.2 video results, longer or higher-resolution LTX-2.5 videos, and audio quality remain unverified. Fewer steps do not reduce base-model memory requirements. See the validation record for known limitations.

Install: Download GenMedia-1.26.1004-arm64.dmg, open it, and drag GenMedia into Applications. The Apple Silicon installer is Developer ID signed, Apple notarized, and verified with Gatekeeper. Model weights are downloaded separately.

Qwen guide · LoRA guide · Full change log (guides in Traditional Chinese)

GenMedia 1.26.1003 build 2312

繁體中文

本次更新讓大量素材、模型與 LoRA 的操作更流暢，維持既有介面、操作方式、生成設定與模型精度。

• 更快瀏覽 Profile：改善大量模型與 LoRA 清單的反應，保留原有排序及安裝狀態。
• 更快找到影片字幕：多部影片共用資料夾時，減少字幕搜尋的重複等待，沿用既有字幕對應方式。
• 更有效率的本機生成流程：加快 GGUF 模型載入時的權重處理與生成進度讀取，減少 LTX 文生影每個步驟的重複準備工作。
• 更清楚的版本識別：版本改用 1.26.1003 build 2312，同日不同 Build 可正常檢查更新。Release Tag 為 v1.26.1003-build.2312。

所有推論維持本機 Swift／MLX，不需安裝 Python；模型權重另行下載。效能改善來自函式量測，不代表整段生成的加速倍率，也不降低大型模型的記憶體需求。實驗性模型及已知 H3 限制請見驗證紀錄。

安裝：下載 GenMedia-1.26.1003-build.2312-arm64.dmg，開啟後將 GenMedia 拖入 Applications。Apple Silicon 安裝包已完成 Developer ID 簽章、Apple 公證及 Gatekeeper 驗證。

已安裝未附 Build 的舊版 1.26.1003 或 1.26.1004 的使用者，請手動下載這次更新，以切換至新的日期加 Build 編號。

完整更新紀錄 · 效能紀錄 · LoRA 使用指南

English

This update makes large asset collections and model/LoRA profiles smoother to work with while preserving the existing interface, controls, generation settings, and model precision.

• Faster Profile browsing: More responsive model and LoRA lists, with the same ordering and installation status.
• Faster video subtitle matching: Less repeated searching when several videos share a folder, while keeping the existing matching rules.
• More efficient local generation workflows: Faster weight processing during GGUF model loading and progress reading, with less repeated preparation at each LTX video generation step.
• Clearer version identification: Versions now use 1.26.1003 build 2312, and update checks recognize newer builds from the same day. The release tag is v1.26.1003-build.2312.

Inference remains local with Swift/MLX; no Python installation is required. Download model weights separately. Improvements are based on function benchmarks and do not represent an overall generation speedup or lower large-model memory requirements. See the validation record for experimental model scope and known H3 limitations.

Install: Download GenMedia-1.26.1003-build.2312-arm64.dmg, open it, and drag GenMedia into Applications. The Apple Silicon installer is Developer ID signed, Apple notarized, and verified with Gatekeeper.

If you use the legacy 1.26.1003 or 1.26.1004 release without a Build suffix, download this update manually once to switch to date-and-build versioning.

Full change log · Performance record · LoRA guide (records and guide in Traditional Chinese)

GenMedia 1.26.1003

繁體中文

新功能與特色

• 4 種新圖像風格：Z-Image Turbo 新增鉛筆素描、吉卜力風格、週六早晨卡通與復古彩色電影 LoRA，從模型中心下載後即可在既有選單切換。
• 4 種影片鏡頭運動：LTX 新增鏡頭推近、拉遠、左移與右移，各附 MLX Q4／GGUF Q3 文生影 Profile。
• 低步數文生影：MiniMax H3 新增 LightX2V 4 步、8 步，以及 Turbo v4（預設 6 步，可調整為 4–8 步）加速 LoRA。對應 Profile 會自動帶入步數與採樣設定。
• 更有效率地使用記憶體：改善原圖預覽與音訊資訊讀取，並減少 Qwen-Image 2.1 生成時的重複運算，維持原有 UI、操作方式及生成設定。
• 延續本機工作流程：沿用模型中心、LoRA 選單與 Profile 編輯器，維持 Swift／MLX 推論，不需另外安裝 Python。

相容性與使用範圍

• LTX 鏡頭 LoRA 原為 LTX-2 19B 訓練，在 LTX-2.3 為實驗性相容。H3 低步數組合亦為實驗性功能；完整影片效果與整段加速倍率尚未驗證。
• LightX2V 支援完整與 Pruned FL2VA GGUF；Turbo v4 限完整 FL2VA。低步數功能目前提供文生影。
• 模型與 LoRA 需另外下載，不包含在安裝包內。減少步數不會降低主模型的記憶體需求。既有 Qwen 2.1 單圖編輯仍屬實驗性功能。

LoRA 使用指南 · 完整更新紀錄

下載

下載 GenMedia 1.26.1003（Apple Silicon）

安裝包已完成 Developer ID 簽章與 Apple 公證，並通過 Gatekeeper 驗證。

SHA-256：f2a9eedf21b34374b0c8565f9873cc3ac733f6b3e8d5705fb86c03634615d5b8

---

English

New features and highlights

• Four new image styles: Z-Image Turbo gains Pencil Sketch, Ghibli Style, Saturday Morning Cartoon, and Technically Color LoRAs. Download them in Model Center and switch styles with the existing selector.
• Four video camera motions: LTX gains dolly in, out, left, and right, each with MLX Q4 / GGUF Q3 text-to-video profiles.
• Low-step video generation: MiniMax H3 gains LightX2V 4-step, LightX2V 8-step, and Turbo v4 acceleration LoRAs. Turbo v4 defaults to 6 steps and supports 4–8; matching profiles apply the step count and sampling settings automatically.
• More efficient memory use: Original-image previews and audio metadata reading use less temporary memory, while Qwen-Image 2.1 reuses work during generation. The existing UI, controls, and generation settings are preserved.
• The same local workflow: Use Model Center, the LoRA selector, and the Profile editor with native Swift/MLX inference and no separate Python installation.

Compatibility and scope

• LTX camera adapters were trained on LTX-2 19B; their use with LTX-2.3 is experimental. H3 low-step combinations are also experimental. Complete video results and end-to-end speedups remain unverified.
• LightX2V supports full and Pruned FL2VA GGUF bases; Turbo v4 requires a full FL2VA base. Low-step profiles currently support text-to-video.
• Models and LoRAs require separate downloads and are not included in the installer. Fewer steps do not reduce base-model memory requirements. Existing Qwen 2.1 single-image editing remains experimental.

LoRA guide · Full update notes (Traditional Chinese)

Download

Download GenMedia 1.26.1003 for Apple Silicon

The installer is signed with Developer ID, notarized by Apple, and verified by Gatekeeper.

SHA-256: f2a9eedf21b34374b0c8565f9873cc3ac733f6b3e8d5705fb86c03634615d5b8

GenMedia 1.26.0922

繁體中文

新功能與操作改善

• 新增 Qwen-Image 2.1：透過提示詞生成圖片，或選取一張圖片進行編輯。文生圖與圖生圖共用同一份 4-bit 模型，從模型中心下載後即可使用；單圖編輯目前為實驗功能。
• Apple Silicon 本機生成：圖片生成與編輯在 Mac 本機執行，不需外部 API 或額外安裝 Python，並可輸出保留透明度的 PNG。
• 連續生成更方便：選取剛完成的圖片後，仍可直接再次文生圖；圖生圖提供獨立按鈕，窄視窗也能完整顯示操作選項。
• 更穩定的工作流程：改善模型下載、暫停、續傳及工作取消，並加強共用素材保護與工作區儲存。
• 更直觀的操作導覽：README 提供操作 GIF，展示圖片、影片、音樂設定及「簡單 MV」流程；同步更新繁中、英文、日文與韓文說明。

使用前須知

• Qwen 2.1 單圖編輯仍屬實驗性。256×256 小尺寸編輯已實測；部分 512×512 編輯曾出現顆粒或金屬感，尚未確認完全解決。
• Qwen 2.1 目前不支援多張參考圖、負面提示詞或 LoRA。MiniMax H3 影片生成也仍處於測試階段。
• 模型需另外下載，不包含在安裝包內。Qwen 2.1 模型適用上游 Qwen Research License。

Qwen 2.1 使用說明 · 完整更新紀錄

下載

下載 GenMedia 1.26.0922（Apple Silicon）

安裝包已完成 Apple 公證與簽章，並通過 Gatekeeper 驗證。

SHA-256：bf2d839ad9803da300f54387f57b42fda9cfef7e74255c8cd5d069f93fa51b03

---

English

New features and improvements

• Qwen-Image 2.1 support: Generate images from prompts or select an image to edit. Text-to-image and image-to-image share one 4-bit model, available in Model Center. Single-image editing is currently experimental.
• Local generation on Apple Silicon: Generate and edit images on your Mac without an external API or a separate Python installation. PNG output preserves transparency.
• Easier repeated generation: Start another text-to-image generation while the previous result remains selected. Image-to-image has its own button, and controls adapt to narrow windows.
• More reliable workflows: Improved model downloads, pause and resume, job cancellation, shared-media protection, and workspace saving.
• Clearer guidance: An animated README walkthrough demonstrates image, video, and music settings and the Simple MV workflow. Documentation is updated in Traditional Chinese, English, Japanese, and Korean.

Before you start

• Qwen 2.1 single-image editing remains experimental. Editing has been tested at 256×256; grainy or metallic-looking results previously seen in some 512×512 edits have not yet been confirmed resolved.
• Qwen 2.1 does not currently support multiple reference images, negative prompts, or LoRA. MiniMax H3 video generation also remains in testing.
• Models require separate downloads and are not included in the installer. Qwen 2.1 model weights are subject to the upstream Qwen Research License.

Qwen 2.1 guide · Full update notes (Traditional Chinese)

Download

Download GenMedia 1.26.0922 for Apple Silicon

The installer is signed, notarized by Apple, and verified by Gatekeeper.

SHA-256: bf2d839ad9803da300f54387f57b42fda9cfef7e74255c8cd5d069f93fa51b03

GenMedia 1.26.0902

Highlights

• Added a MiniMax H3 low-resolution quality warning in the video generation UI. The dialog shows the requested and aligned dimensions and lets the user cancel or continue.
• MiniMax H3 enforces a technical minimum of 64×64 and aligns dimensions down to a 32-pixel grid. A short side below 512 px now requires confirmation; the native quality reference is approximately 768 px.
• Fixed the MiniMax H3 video writer back-pressure path so long videos with audio do not stall while submitting media samples.

H3 Status

MiniMax H3 remains in the testing phase. Generation results and all H3 GGUF variants are not guaranteed to be correct across different hardware configurations.

Signed Artifact

• GenMedia-1.26.0902-arm64.dmg
• Developer ID signed app with a stapled notarization ticket
• SHA-256: ed5e99557101317e105189ca8b4e09090cd2d66db58dbf3da932bedb211a7b30

GenMedia 1.26.0901

Highlights

• Fixed a MiniMax H3 latent reshape crash caused by dimensions such as 1280×720. The app now aligns H3 width and height down to the required 32-pixel grid (1280×720 becomes 1280×704), and the Worker rejects invalid geometry before patchification.
• Added an explicit JSON request path for the H3 Worker while preserving explicit --format and --variant dispatch. The Worker does not guess model formats from directory contents.
• Added the H3 Metal-side GGUF quantization path. The CPU reference quantizer remains available for diagnostics.
• Updated H3 profiles to use a safe default height of 704 pixels.
• Added H3 keyframe Video VAE encoding support for the connected conditioning path.

Verification

• swift build -c release — passed.
• swift build --package-path RuntimeSupport/MiniMaxH3Worker -c release — passed.
• MiniMax H3 geometry tests — 3/3 passed.
• Real MiniMax H3 GGUF generation — passed at 640×384, 4 frames, with audio.
• codesign --verify --deep --strict --verbose=2 dist/GenMedia.app — passed.

Important

• MiniMax H3 remains in a testing phase. Correctness is not guaranteed for every H3 GGUF variant, generated result, or hardware configuration.
• The Qwen3-VL vision tower for Ref2VA, long/high-resolution Video VAE temporal and spatial tiling, and multi-frame clip anchors remain outside the stable H3 scope.
• The application has removed most Python dependencies. Complete Python removal remains the long-term goal, and remaining issues will continue to be fixed.
• The full H3 Worker suite still contains two known temporal-tiling test gaps; they are not represented as passing in this release.

Download

• DMG: GenMedia-1.26.0901-arm64.dmg
• SHA-256: eb4b42e5100a01af5492a2c78c80eb11d74c8002eb6a4e91771ceabea19023d4

GenMedia 1.26.0831

Highlights

• Native Swift/MLX runtime updates for MiniMax H3, LTX, GGUF, subtitle translation, and Profile management.
• Consolidated the LTX MLX and GGUF workers into one shared runtime, reducing duplicated runtime packaging while preserving explicit MLX/GGUF dispatch.
• GenImageDoctor remains available as a CLI target but is no longer shipped inside the application bundle, further reducing the packaged footprint.
• Shared MLX metallib packaging and signed Worker fixes.
• Most Python dependencies have been removed from the application; complete Python removal remains the long-term goal.

Important

• MiniMax H3 is currently experimental/testing. Correctness is not guaranteed for every model variant, output, or hardware configuration.
• If H3 output is incorrect, retain the selected Profile, model variant, and Runtime log for follow-up fixes.

Distribution

• DMG: GenMedia-1.26.0831-arm64.dmg
• SHA-256: 3ea141e96ed872e6be9c8b49ff146650ec707823aa1579cab00e8513b419d98d

GenMedia 1.26.0830

GenMedia 1.26.0830

What's New

• Expanded native Swift/Core ML inference paths for LTX, GGUF, subtitles, and Whisper Small.
• Added Whisper Small as the speed-focused recommended model while retaining Whisper Large v3 Turbo.
• Improved subtitle original-text preservation, subtitle sidecar matching, player subtitle loading, and generation progress reporting.
• Updated model downloads, Runtime dispatch, the isolated Z-Image Worker, and MLX metallib version management.

Python Runtime Migration

The system has removed most Python dependencies, with the final goal of removing Python completely. If some features encounter issues after the migration, they will continue to be fixed as native Swift Runtimes are completed. This release focuses on replacing the legacy Python paths with native implementations, while some edge cases may still require follow-up fixes.

Download

• Apple Silicon DMG: GenMedia-1.26.0830-arm64.dmg
• SHA-256: e24ae7879225af82d865495e12d7f74bfefdfcc23bd8498b18ab5759961518b2

GenMedia 1.26.0827

Native Apple Silicon release for macOS 14 or later.

Highlights

• Added Civitai API Token settings backed by macOS Keychain; Civitai LoRAs now download directly with HTTPS Bearer authorization instead of opening the website for manual download.
• Replaced the MiniMax Music 3 app path with the bundled Swift worker, including real frame, denoising, vocoder progress events and cancellation support.
• Added the embedded LGPL FFmpeg compatibility layer for broader video/audio import, playback proxies, remuxing, conversion, image loops and media merging.
• Added automatic workflow templates with workspace and tab preparation for Simple MV, image looping and media merging.
• Retained multilingual subtitle generation, multimodal model profiles, workspace persistence and the existing MCP integrations.

Download

The Apple Silicon DMG is signed with Developer ID, notarized by Apple, and stapled for offline Gatekeeper verification.

Build: 1847

SHA-256: 59dbecab7b812e17a709019c679f516bf7d9cfc668e47a7a0c3771ea796dbd46

GenMedia 1.26.0826

Native Apple Silicon release for macOS 14 or later.

Highlights

• Added native subtitle generation from video and audio with multilingual Whisper, Chinese Paraformer, and Japanese Parakeet profiles, with SRT/WebVTT export and optional local Qwen translation.
• Expanded subtitle translation to 25 target languages across the app and standalone MCP server.
• Subtitle files now default to the source media folder and reuse the source base filename.
• Added safe media removal choices for keeping or deleting the underlying file.
• Updated MiniMax Music 3 to use 20 inference steps by default.
• Refined workspace deletion styling and keeps the MCP API panel visible in a disabled state while the service is off.
• Added multimodal Qwen3-VL, Qwen3.5, and Qwen3.8 profiles for both image-to-text and text-to-text workflows, including complete multimodal model-file verification.
• Added named workspaces with independent project tabs, persistent assets, and resumable lineage.

Download

The Apple Silicon DMG is signed with Developer ID, notarized by Apple, and stapled for offline Gatekeeper verification.

Build: 1711

SHA-256: 46f7aa3fe5a8814140fc19612667883a973e6aab7ac34c4a01dc06f4228089af