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)