每日生態雷達 · 2026-11-21RSS Feed

H3 模型動態與更新日誌

按時間軸彙整官方模型發佈、權威榜單異動、高價值工作流發現與 H3Gen 平台升級紀錄,每條外部動態皆附真實來源連結。

  1. H3Gen 平台更新H3Gen v2.4

    Full-Duration Protected Preview Player, FL2VA Keyframe Studio & H3-Image 4K Launch

    H3Gen rolled out full-duration protected video previews, a dedicated First & Last Frame keyframe studio, the Weekly Viral Prompts hub, Creator Playbooks, and MiniMax-H3-Image 4K generation.

    • Unpaid trial renders now play 100% of their full clip duration in-page with blob-stream protection instead of truncating at 2 seconds.
    • Launched /first-last-frame-video with dual-card FL2VA keyframe highlighting across all 8 locales.
    • Added /h3-image studio with 1–16 reference image support and 1-click handoff into H3 Video.
    H3Gen Release Notes立即體驗
  2. H3Gen 平台更新H3Gen v2.3

    16 One-Click Audio-Video Templates & Native Mobile Bottom-Sheet Studio

    Shipped 16 production-ready commercial, UGC, kinetic poster and cinema templates with dedicated 384×384 product sample visuals and a mobile bottom-sheet modal.

    • 1-click sample product photo loader + balanced [Sample → Upload Photo] dropzone across 16 templates.
    • 2-column mobile masonry template feed and sticky bottom generation bar safe for iOS Chrome & Safari.
    H3Gen Release Notes立即體驗
  3. 每日玩法雷達

    Workflow Radar: 3-Part Role + Beat + Stereo Foley Prompting Outperforms Dense Paragraphs

    Community evaluations across multi-reference runs show significantly stronger identity lock and lip-sync timing when prompts explicitly separate asset roles, timestamped camera beats, and acoustic cues.

    • Name uploaded assets explicitly (Image 1 = subject identity, Video 1 = camera path, Audio 1 = vocal cadence).
    • Timestamp major motion transitions (0–3s, 3–6s, 6–8s) for cleaner 2K commercial reveals.
  4. 榜單與生態進展Ecosystem

    MiniMax-H3-Image Tops Open-Weight AA-Image v2.1 Editing & T2I Leaderboards

    Independent blind evaluations on Artificial Analysis placed MiniMax-H3-Image #1 among open-weight models for both text-to-image and multi-reference image editing.

    • Direct 6-step native 4K (~16 MP) synthesis without requiring a secondary upscaler pass.
    • Supports 1 to 16 reference images for multi-subject composition and product background replacement.
  5. 官方模型發佈H3-Image

    MiniMax Releases H3-Image & H3-Image-Highspeed Multimodal Image Foundation Model

    Derived from the H3 multimodal transformer architecture, H3-Image brings native 4K generation and up to 16-image reference editing into the H3 family.

    • Ideal upstream companion for H3 Video: generate clean packshots or character sheets, then animate with stereo sound.
    • Available in standard high-fidelity and Highspeed low-latency variants.
  6. 榜單與生態進展Video Arena

    MiniMax H3 Climbs to Top-3 Global Standing in Audio-Video Blind Arena

    Blind human preference benchmarks ranked MiniMax H3 #1 in Video Editing with Audio, #2 in Text-to-Video with Audio, and #3 in Image-to-Video with Audio.

    • Single-pass 32 kHz stereo synthesis eliminates drift between spoken dialogue, foley impacts and camera depth.
    • Strong edge preservation on kinetic poster headlines and commercial product geometry.
  7. 官方模型發佈FL2VA / Ref2VA

    H3-Base Checkpoints (FL2VA & Ref2VA) Document Dual Keyframe & Omni-Reference Paths

    Technical documentation detailed H3’s specialized conditioning branches: FL2VA for exact start/end frame anchoring and Ref2VA for multimodal asset fusion.

    • FL2VA optimizes visual and acoustic fidelity across single-frame and first-plus-last-frame transitions.
    • Ref2VA unifies up to 9 images, 3 reference videos and 3 audio clips inside a shared transformer context.
  8. 官方模型發佈MiniMax H3

    MiniMax H3 Omni-Modal Video-Audio Foundation Model Debuts

    MiniMax unveiled H3, a unified multimodal video generation model producing native 2K 24fps video and synchronized stereo audio in a single pass.

    • Supports text, first/last frames, reference images, motion reference videos and voice/music audio inputs.
    • Renders 4–15s native 2K clips and extended single-take sequences with coherent room acoustics.