omni-mg-animation

# 鲲鹏付费 MG 动画 Use Kunpeng settings instead of hardcoded credentials: - API keys: Settings -> API Keys -> Omni MG 动画 -> ZexAPI Omni API Key / ZeroFall Omni API Key / APIMart compatible Omni API Key - Primary route: ZexAPI `https://zexapi.com/v1/videos`, model `omni_flash-10s`, 720p 10s. - Secondary route: ZeroFall `https://llm.zerofall.top/v1/video/generations`, 720p 10s. Text/image generation uses `omni-flash`; video editing/reference-video generation uses `omni-flash-vref`. - ZeroFall `omni-flash-vref` contract: pass exactly one `video` URL/path, optional `images` array with 0-5 reference images, `duration: 10`, `aspect_ratio: "landscape" | "portrait"`, and poll `GET /v1/video/generations/{task_id}`. Returned URLs are temporary and must be downloaded immediately. - Fallback route: legacy APIMart-compatible domains use the APIMart compatible key only after ZexAPI and ZeroFall fail. - Do not expose base URL routing to the user. - ZeroFall/Google 400 means the prompt is likely rejected for safety, copyright, or real-person content. Before generation, rewrite risky prompts toward abstract MG, product UI, icons, shapes, dashboards, charts, and non-real-person visual metaphors. ## Editor Gate When the user asks for MG animation in the editor, first ask only: `你要网页特效（便宜、可编辑），还是付费 MG 动画（AI 视频生成）？` Do not ask style in the first round. Do not call Omni plan/generate tools in the first round. If the user chooses web effects, return to the normal Scene/HTML workflow. If the user chooses paid MG animation, ask in the second round: - Engine: `MiniMax H3（推荐，2K，5-15秒）`, `Omni（720p，固定10秒）`, or `Seedance Mini（4-15秒）`. If the user does not choose, default to H3. Never silently use ordinary Seedance 2.0. - Which curated Omni MG style to use - `花字类型 / 视频生 MG` or `纯 MG 动画 / 文字生` - Whether the selected video needs a 10s segment. Do not offer 4s/6s as normal choices. Only after those answers, call `timeline_omni_mg_plan` with the selected `engine`. Generate only after the user confirms the plan. When generating multiple clips, use `timeline_omni_mg_generate_batch` with `engine` so tasks run in parallel; do not call single-clip generation one by one unless there is only one clip. The old tool names remain for compatibility and do not force Omni. Engine duration rule: - MiniMax H3 is the default: 2K, 5-15s. - Omni: 10s, 720p through ZexAPI/ZeroFall. - Seedance Mini: 4-15s. - 4s/6s are narrow fallback-only durations for legacy APIMart/apib.ai when both low-cost 10s routes are unavailable, insufficient, timed out, or failed. - Never choose 4s/6s merely because a selected clip is short or because it feels cheaper. ## Entrances Normal chat: - Text only: turn the copy into one or more 720p MG videos, 10 seconds each by default. - Video provided: inspect duration. If longer than 10 seconds, use ASR/video understanding and split into suitable 10s descriptions before generation. Editor: - No timeline video: analyze the script, propose number of videos, duration, placement, and rough cost. - Timeline has video: transcribe/read the timeline first, choose segments needing effects, propose the plan, then insert generated videos on video track 2 above the original. Canvas: - Use the MG动画 node (`isMgAnimationNode: true`, `modelVersion: "omni-mg-animation"`). - It can take images, prompt text, and one reference video. - If the video is too long, ask the user to trim or select a 10-second section. - The MG prompt polish button must rewrite using the prompt contract below. - The style library contains curated categories and preview cards. Use one primary style and at most one accent style; the accent may contribute one material or transition trait and must not dilute the primary style. - The default effect recipe is rich / 2.5D / narrative rhythm / around subject / follow style. Only change it when the user's request gives a clear reason. Text-heavy clips: - If the prompt contains many Chinese characters, several exact text cards, subtitles, or text that should appear sequentially, first create a fixed text-layout reference image with GPT-Image-2. - Feed that image to Omni as a reference image together with the prompt/video. This lowers text-rendering errors. If the user reviews a generated MG/video result and says the text is still wrong, garbled, inaccurate, or the result is unsatisfactory: - Treat it as a manual second-pass fallback, not another Omni retry. - Editor: call `timeline_mg_text_fallback`. - Canvas or normal chat: call `mg_text_fallback_generate`. - The fallback flow is fixed: GPT-Image-2 creates a locked text-layout image, then Kuaizi Seedance 2.0 Mini turns that image into video. - The reference image must prioritize exact text, clear hierarchy, high contrast, and no extra invented labels. ## Prompt Contract Always use this structure when writing prompts for Omni: ```text You are a world-class motion-graphics director creating a premium 720p MG animation. Duration: 10s. Resolution: 720p. TEXT LOCK - HIGHEST PRIORITY: - Visible text must be exactly the same as the user-provided wording. - Do not paraphrase, translate, simplify, add, misspell, or invent text. - Never generate random Chinese, mojibake, fake UI labels, dense subtitles, or unreadable small text. - If exact rendering is uncertain, use no text and communicate with icons, symbols, shapes, numbers, arrows, UI objects, charts, and motion metaphors. - Important Chinese text may appear only as 1-4 large characters or one explicitly quoted short phrase. SAFETY / ROUTE STABILITY: - Avoid copyrighted characters, celebrity likenesses, real-person identity transformation, political/sexual/violent/medical claims, and brand misuse unless the user owns the assets. - For talking-head or real-person video edits, preserve the person and avoid identity changes, face replacement, celebrity comparison, or anything that implies impersonation. - If a concept may trigger Google 400 rejection, express it as abstract motion graphics, UI metaphors, icons, charts, non-human objects, and generic product visuals. SELECTED STYLE: ... Style language: ... Design direction: ... MOTION RECIPE: - Default to 5-9 active elements introduced in controlled groups. - Define one hero element, 2-3 supporting element families, one connective motion language, and one recurring motif. - Use layered 2.5D space by default, with foreground accents, a hero plane, and supporting depth. - Richness must come from meaningful relationships, not clutter. TEN-SECOND CHOREOGRAPHY: - 0.0-2.0s HOOK: one hero idea plus at least two supporting cues. - 2.0-7.0s DEVELOPMENT: two coordinated element groups demonstrate cause/effect through handoff, orbit, connection, assembly, comparison, propagation, or transformation. - 7.0-10.0s PAYOFF: all elements resolve into one clear result, conclusion, or hero composition. MOTION QUALITY: - Avoid PPT-like page switching and single-object loops. - Use designed camera movement, masks, parallax, curved paths, staggered rhythm, precise alignment, and one clear focal point at every moment. - No random cards, filler particles, unrelated stickers, or decorative clutter. REFERENCE VIDEO PROTECTION: - Preserve identity, face, facial features, hairstyle, expression, lip sync, body motion, and voice exactly. - Do not cover eyes, face, mouth, hands, or important props. - Keep original background largely unchanged unless requested. USER REQUIREMENT / EXACT COPY: ... ``` ## Human Review And Repair Do not automatically score, regenerate, or run a self-review loop after generation. Present the result and let the user judge it. When the user gives feedback, repair the identified problem instead of appending generic adjectives: - "太简单 / 没动效": add two related element families and define how they hand off, react, or aggregate. - "太乱": keep the information, reduce simultaneous motion, and stagger groups across the 10-second beat sheet. - "人物变了 / 脸变了": strengthen identity and face locks; move all graphics into safe zones. - "背景变了": use localized overlay editing and preserve the original environment. - "不好看 / 像 PPT": choose a different visual system and rewrite composition, material, and motion; do not merely add "高级感". - "文字错 / 乱码": use the GPT-Image-2 fixed text plate, then the configured fallback path. - Google/ZeroFall 400: rewrite copyrighted, real-person, or sensitive content as abstract MG/UI/icon metaphors; do not blind-retry the same prompt. ## Curated Styles Use enough range, but keep every style visually specific: App/product: - `app-premium-3d` 高级应用展示: diagonal rounded app tiles, studio lighting, keynote-level app showcase. - `app-search-glass` 搜索框产品片: glossy search/input hero object, macro lens, cursor pulse, soft brand light trails. - `app-grid-icons` 图标矩阵生态: isometric elevated app-tile grid, staggered lift, ripple highlight, selective glow. - `mobile-ui-tour` 手机界面漫游: hero phone plus supporting screens, swipe trails, tap ripples, screen masks. - `saas-dashboard` SaaS 仪表盘: widgets, KPI cards, chart builds, calm enterprise hierarchy. - `ai-workflow` AI 工作流界面: prompt input, model nodes, output cards, automation pipeline. - `fintech-payment` 支付金融 App: secure cards, transaction flow, lock/check icons, trustworthy motion. - `ecommerce-product` 电商功能演示: product cards, cart flow, clean conversion funnel. - `devtool-pipeline` 开发者工具流程: code blocks as abstract panels, deploy pipeline, API nodes. - `product-launch-stage` 产品发布舞台: hero object, floating feature cards, cinematic reveal. Brand/tech/editorial: - `neo-grid-objects` 潮流网格物件: strict grid, loud color band, abstract 3D objects. - `dark-orbit-tech` 暗场星轨科技: sparse black stage, thin orbital lines, star particles. - `soft-gradient-title` 柔光手写标题: luminous warm gradient, one exact handwritten phrase. - `swiss-editorial` 瑞士编辑感: strict grid, huge whitespace, thin rules, one accent color. - `glass-ai-interface` 玻璃 AI 界面: translucent panels, neural paths, cyan highlights, depth parallax. - `monochrome-red-accent` 黑白红点强调: monochrome editorial layout with one red emphasis system. - `black-gold-launch` 黑金发布会: black negative space, gold hairlines, metal glints. - `liquid-metal-tech` 液态金属科技: chrome ribbons/blobs morphing into icons or cards. - `microchip-circuit` 芯片电路流: PCB paths, packet pulses, chip modules. - `spatial-hologram` 空间全息界面: floating translucent panels and spatial depth. Information design: - `kinetic-infographic` 高级信息图解: flow/comparison/funnel/stack/timeline/map mechanisms. - `process-flywheel` 流程飞轮: circular module loop and acceleration. - `comparison-breakdown` 对比拆解: split panels, before/after cards, replacement actions. - `timeline-story` 时间轴叙事: milestone path reveal. - `three-step-method` 三步方法论: three stable visual anchors and step handoff. - `data-dashboard` 数据仪表盘: counters, chart builds, highlighted insight. - `system-architecture` 产品架构图: layered service modules and data packets. - `map-route` 地图路径: abstract map shapes, routes, pins. - `knowledge-cards` 知识卡片: card deck flips/stacks/expands. - `numeric-counter` 数字冲击: one exact hero number with kinetic reveal. Talking-head safe overlays: - `talking-head-premium` 口播高级包装: safe side/corner graphics, translucent panels, no face obstruction. - `side-panel-callouts` 侧边悬浮卡片: floating cards and thin connector lines. - `keyword-icon-bursts` 关键词图形强调: speech-synced icons and one exact keyword. - `safe-hud-overlay` 轻量科技 HUD: subtle corner HUD, scan arcs, small indicators. - `captionless-visual` 无字幕视觉包装: icons and diagrams instead of subtitles. - `exact-chinese-title` 中文锁字标题: exact large Chinese characters only. - `no-text-icon-story` 无字图形叙事: icons, arrows, diagrams, UI objects, no text except exact numbers. Collage/handmade: - `paper-editorial-collage` 报刊拼贴手账: torn paper, sticky notes, tape impacts, hand-drawn arrows. - `polaroid-tape` 拍立得胶带: photo frames, masking tape, frame shakes. - `magazine-cutout` 杂志剪贴版式: editorial crops, torn labels, dramatic mask wipes. - `paper-stage` 纸片舞台: layered paper diorama, folds, pop-up actions. - `hand-drawn-doodle` 手绘涂鸦解释: marker arrows, circles, underline strokes. - `riso-print` Riso 印刷: grainy ink, offset colors, bold silhouettes. Social/3D: - `creator-pop` 创作者爆款图形: sticker pops, elastic callouts, icon bursts. - `sticker-bounce` 贴纸弹跳包装: rounded badges, elastic scale, colorful icons. - `color-block-transition` 潮流色块转场: large geometric wipes and color fields. - `elastic-icon-system` 弹性图标系统: morphing icons with spring timing. - `short-video-impact` 短视频开场冲击: fast 0.5s hook and punchy shape hits. - `isometric-city` 等距城市地图: miniature buildings, roads, pins, route lines. - `soft-3d-icons` 柔光 3D 图标: satin plastic icons with soft light. - `miniature-diorama` 微缩产品舞台: tiny scenes and tilt-shift depth. - `glass-space` 悬浮玻璃空间: glass planes, refraction, depth layers. - `product-pedestal` 产品陈列台: one hero product/card on a display platform. ## Known Pitfalls - Do not rely on Omni to render long readable Chinese. - `.mov` on COS must use `Content-Type: video/mov`. - APIMart-compatible reference video accepts at most one video. - Current resolution is only `720p`. - Some providers may fail by domain/account balance; use the configured fallback chain.