[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7DAV3WSENYRMM624Y2TXMMJ":3,"skill-same:01M7DAV3WSENYRMM624Y2TXMMJ":55},{"skillNo":4,"name":5,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":23,"commitSha":24,"skillDirectory":25,"forksCount":26,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":6,"ownerDescription":29,"updatedAt":30,"install":31,"resources":44},"01M7DAV3WSENYRMM624Y2TXMMJ","higgsfield-seedance-2-5",null,"alicewe1\u002Falice_skill","alicewe1","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F102134151?v=4","01M7D9STCM8HAESM958JP0989K","Alice_we","01M7D9STDR22MTD9FTTWPAZNV8",64,"GPL-3.0","https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill\u002Fblob\u002Fmain\u002F_modules\u002Fhiggsfield-seedance-2-5\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: higgsfield-seedance-2-5\ndescription: \"Seedance 2.5 prompt director — the omni-reference dialect. Routes the four generation modes (t2v \u002F omni_reference \u002F video_edit \u002F video_extension), writes explicit @Image\u002F@Video\u002F@Audio reference roles with exclusions, stages 30-second videos into end-state beats, and covers video editing, forward\u002Fbackward extension, first-last-frame and multi-keyframe control, storyboard grids, blockout rendering, and seamless transitions. Use whenever the user asks for a Seedance 2.5 prompt, mentions Seedance 2.5 \u002F Dreamina \u002F Jimeng, wants a clip longer than 15s on Seedance, wants to EDIT or EXTEND an existing video rather than generate a new one, or supplies more than a handful of image\u002Fvideo\u002Faudio references. For Seedance 2.0 (4K, start\u002Fend frames, genre hint) use higgsfield-seedance instead.\"\nuser-invocable: true\nmetadata:\n  tags: [higgsfield, seedance, seedance-2.5, dreamina, jimeng, omni-reference, video-edit, video-extension, multi-reference, long-video, keyframes, storyboard, blockout, transitions]\n  version: 1.2.0\n  updated: 2026-08-09\n  parent: higgsfield\nx-alice-class: assist\n---# Higgsfield Seedance 2.5 Director\n\nSeedance 2.5 is a **different dialect from Seedance 2.0**, not a version bump you can\nprompt through by habit. 2.0 is a reference-driven shot generator with start\u002Fend frames\nand a 4K lane. 2.5 is an **omni-reference production model**: up to 50 reference\nmaterials, 30-second native runtime, and three non-generation modes — it can edit a video\nyou already have, and extend one forward or backward from its boundary frame.\n\nThe prompt grammar changes with it. Reference roles are declared in prose (`@Image 1\ndefines …`), audio and text get bracket syntax, long videos are staged with explicit end\nstates, and first\u002Flast frames are announced **inside the prompt** rather than selected as\na mode.\n\n> **Model split — read this before writing anything.** 2.5 caps at **720p** and has no\n> `start_image` \u002F `end_image` media role, no `genre` hint, and no 4K lane. If the job\n> needs 4K, a genre hint, or platform-level start\u002Fend frame pinning, it is a\n> **Seedance 2.0** job — `..\u002Fhiggsfield-seedance\u002FSKILL.md`. See § Choosing 2.0 vs 2.5.\n\n## QUICK FACTS\n*Generated-checked block (scripts\u002Fbuild_index.py verifies anchors). Routing aids — read the linked sections for the rules themselves.*\n- Four modes, picked **before** writing: `t2v` · `omni_reference` · `video_edit` · `video_extension`; the mode changes what the prompt *is* [→](#the-mode-router)\n- Higgsfield surface: **480p\u002F720p only**, duration **4–30s**, no start\u002Fend-frame role, no genre hint, `extension_mode` required for (and only for) `video_extension` [→](#the-higgsfield-parameter-surface)\n- `video_edit` **ignores** `duration` and `aspect_ratio` and bills by the source video's length; `video_extension` inherits the source's aspect ratio [→](#the-higgsfield-parameter-surface)\n- Every reference material gets an explicit role **and** an exclusion — \"what to use\" plus \"what not to use\"; never let the model infer the mapping [→](#reference-roles-say-what-to-use-and-what-not-to-use)\n- Each material also declares a **fidelity grade** — full-preserve \u002F partial-preserve \u002F attribute-transfer (name the target) \u002F loose-guide; beat lines name characters (name + one visible marker), never handles [→](#fidelity-say-how-much-of-each-material-must-survive)\n- Material budget: 30 images \u002F 10 videos ≤30s total \u002F 10 audio ≤30s total, 50 materials max; stability ranges are 1–8 subjects (images), 1–5 subjects at 5–10s (video\u002Faudio) [→](#material-budget)\n- Multi-reference is a 5-step workflow — map → group → profile → select-by-scene, one line per subject; `@Images 1 through 4 define four characters` is the canonical failure [→](#multi-reference-the-five-step-workflow)\n- Long videos are **staged**, not paragraphed: one primary change per stage + an explicit **end state**; timestamps allocate a budget, they are not frame-accurate edit points [→](#long-video-stages-and-end-states)\n- Bracket syntax: `()` music · `\u003C>` SFX · `{}` dialogue · `【】` subtitles; non-Chinese dialogue needs a language line before the line [→](#audio-and-text-bracket-syntax)\n- First\u002Flast frames and multi-keyframes are declared **in the prompt** (`@Image 1 is the first frame`) — aspect ratio locks to the first image; never merge the two anchors into one sentence [→](#first-last-frame-and-multi-keyframe-control)\n- Editing needs a **sole editing master** + edit scope + Timeline Inheritance; extension needs the **boundary frame aligned before** any new content: `MODE-PLAYBOOKS.md`\n- Storyboard grids, coarse-vs-fine blockouts, one-click video, seamless transitions: `MODE-PLAYBOOKS.md`\n- **AI-VFX production pipeline** — model-per-asset-class routing, the size-ref frame, location batching, the `omni_reference` v2v lane (source ≥4s, duration = source), the four-batch rule, the slop catalog: `VFX-PIPELINE.md`\n- Emotion needs 2–4 **observable** cues, not adjectives; niche camera terms get translated into a visible result [→](#emotional-direction-and-camera-terms)\n- The real-person formula is 7 slots — and slot 1 is **role, never age**: the age-blind engine rule outranks the source guide's `[Age\u002FRace]` label [→](#the-real-person-character-formula)\n- Hard limits that must not be over-promised (frame accuracy, locked parameters, pixel-identical transitions) [→](#hard-limits-do-not-over-promise-these)\n- Dreamina-product features that are **not** on the Higgsfield surface — Ultra Long Video 180s, mark-based editing, Clay Renderer [→](#dreamina-only-what-higgsfield-does-not-expose)\n\n---\n\n## Provenance\n\nTwo independent sources, labelled throughout:\n\n| Label | Source |\n|---|---|\n| `[OFFICIAL — Dreamina]` | ByteDance's *Dreamina Seedance 2.5 Prompt Guide* + *User Guide* — the model vendor's own prompt doctrine. Prompt grammar is model-side, so it carries across to Higgsfield's hosting. |\n| `[OFFICIAL — platform]` | Higgsfield's live `models_explore` catalog, snapshot **2026-08-07** (`..\u002F..\u002Fspecs\u002Fmodel-specs.json`). Parameters, enums, and media roles come from here and nowhere else. |\n| `[DREAMINA-ONLY]` | A Dreamina *product* feature with no Higgsfield parameter behind it. Never quote these as things the user can do here. |\n\nWhere the two disagree about what is *settable*, the platform snapshot wins — it is what\nthe API actually accepts.\n\n---\n\n## The Mode Router\n\nPick the mode first. The same sentence means different things in different modes, and two\nof the four modes are not generation at all.\n\n| The user wants | Mode | What the prompt is |\n|---|---|---|\n| A clip from a description, no materials | `t2v` | A scene brief — the core formula below |\n| A clip built from images \u002F videos \u002F audio they supply | `omni_reference` | A **role map** plus a scene brief |\n| To change something inside a video they already have | `video_edit` | An **edit order**: master + scope + preserve list |\n| More footage before or after a video they already have | `video_extension` | A **boundary contract** plus new content |\n\nTwo rules that fall out of this:\n\n1. **First\u002Flast frames, keyframes, storyboard grids, and blockouts are all `omni_reference`.**\n   2.5 has no separate first\u002Flast-frame mode on this platform — the anchor images are\n   ordinary references whose *role sentence* says they are the first and last frame.\n   `[OFFICIAL — Dreamina: \"no need to switch to a separate first\u002Flast-frame mode\"]`\n2. **Editing is not regeneration.** If the user wants the shot rebuilt, that is\n   `omni_reference` with the old clip as a motion reference — not `video_edit`. `video_edit`\n   preserves the master's timeline and changes one scoped thing inside it.\n\n> **Video-to-video is not automatically `video_edit`.** The field VFX workflow — swap the\n> person in this plate, keep every other pixel — runs in **`omni_reference` with the source\n> attached as a video reference**, because that is the lane where `duration` is settable and\n> must be **matched to the source** (and where the source therefore has to be **≥ 4 s**, the\n> `duration` floor). `video_edit` is the lane for a scoped change inside a master whose\n> timeline must survive untouched. Full routing table + the performance-inheritance clause:\n> `VFX-PIPELINE.md` § Stage 4. `[FIELD — AI-vs-VFX, 2026-08-08]`\n\n---\n\n## The Higgsfield Parameter Surface\n\n`[OFFICIAL — platform, snapshot 2026-08-07]` · verify against `..\u002F..\u002Fspecs\u002Fmodel-specs.json`\nbefore quoting (HARD RULE 3).\n\n| Parameter | Values | Notes |\n|---|---|---|\n| `mode` | `t2v` · `omni_reference` · `video_edit` · `video_extension` | default `t2v` |\n| `duration` | 4–30 s | default 5 — **ignored in `video_edit`** |\n| `resolution` | `480p` · `720p` | default 720p — **there is no 1080p or 4K on 2.5** |\n| `generate_audio` | bool | default true |\n| `extension_mode` | `backward` · `forward` | **required** for `video_extension`, **not allowed** otherwise |\n| aspect ratio | `auto` · `21:9` · `16:9` · `4:3` · `1:1` · `3:4` · `9:16` | ignored in `video_edit`; follows the source in `video_extension` |\n| media roles | `image_references` · `video_references` · `audio_references` | **no `start_image` \u002F `end_image`** |\n\nThree consequences worth stating to the user before they spend credits:\n\n- **`video_edit` bills by the source video's duration**, and neither `duration` nor\n  `aspect_ratio` is settable — a 20-second master costs a 20-second render no matter how\n  small the edit.\n- **`video_extension` inherits the source's aspect ratio**; only the extension's\n  *duration* is yours to set.\n- **No `genre` parameter.** 2.0's genre hint does not exist here — genre lives in the\n  prompt's visual-style clause instead.\n\nPreflight the same way as 2.0:\n\n```\npython3 scripts\u002Fseedance_lint.py --preflight --model seedance_2_5 \"\u003Cprompt>\"\n```\n\nThe linter reads the enums out of `..\u002F..\u002Fspecs\u002Fmodel-specs.json`, so an out-of-range duration, a\n1080p request, or a `video_extension` missing its `extension_mode` is caught before the\nrender.\n\n---\n\n## The Core Prompt Formula\n\n`[OFFICIAL — Dreamina]` Combine only the parts the shot needs; omit the rest rather than\npadding empty slots.\n\n```\n\u003CSubject> performs \u003Cprimary action or event> in \u003Cscene and environment>.\nThe visuals feature \u003Cvisual style>.\nUse \u003Cshot size, camera angle, camera movement, or cuts>.\nAudio includes \u003Cdialogue, ambience, sound effects, or music>.\n```\n\n- **Subject + action** is load-bearing — make it concrete. \"The man runs\" → \"the man\n  accelerates into a sprint while his jacket reacts to the airflow.\"\n- **Scene and environment** — location, time, weather, spatial relationships, background state.\n- **Visual style** — lighting, color, materials, texture, mood. Only descriptors that add\n  information; stacked buzzwords (\"cinematic, 8K, masterpiece\") sample nothing in particular.\n  The named-substitute discipline in `..\u002Fhiggsfield-seedance\u002FSKILL.md` § Prompt-Craft Laws\n  applies unchanged.\n- **Camera** — shot size, angle, movement, focus subject, transitions. Motion matches the\n  action; it is not decoration.\n- **Audio** — dialogue, voice characteristics, ambience, SFX, music, synchronized to the visuals.\n\n**Generation parameters are not prompt text.** Resolution, duration, and aspect ratio are\nset on the generation page or via the API — writing them into the prose does nothing except\nin the modes that auto-lock them, where they are not settable at all.\n\n---\n\n## Reference Roles — Say What to Use *and* What Not to Use\n\nThe moment there is more than one material, or a material sitting next to a text\ndescription, the prompt must state what each material contributes. `[OFFICIAL — Dreamina]`\n\n```\n@Image 1 defines \u003Csubject>'s \u003Cappearance, clothing, structure, or material>.\n@Video 1 defines \u003Cmotion, camera movement, or pacing>.\n@Audio 1 defines \u003Ccharacter or sound type>'s \u003Cvoice, dialogue, ambience, or music>.\n```\n\nEvery material that could leak something unwanted gets an explicit exclusion in the same\nsentence:\n\n```\n@Image 2 defines the workbench and window light. Do not use the people in the image.\n```\n\nRules:\n\n- **Mappings live in the prompt.** Text labels drawn inside an image are not a mapping, and\n  the model will not infer which person or prop a material represents.\n- **Video-only references are motion\u002Fpacing references by default**, not identity — say so\n  explicitly when you mean otherwise.\n- **Several views of one subject must say they are one subject**: \"All four images define\n  one folding desk lamp. The output must contain only one lamp throughout.\" Without it the\n  model duplicates the subject.\n- **When a reference video already carries the motion accurately, state only which\n  attributes to inherit.** Restating every action fights the reference. A blockout or motion\n  video carries motion and spatial structure — not identity — so the prompt still has to\n  define subjects, scene, action, and visual style.\n- **Never place a reference handle in a shot where that subject is absent** — the same rule\n  as 2.0's tag discipline; the model forces it into frame.\n- **Character sheets leak their staging.** A sheet's neutral backdrop and multi-view panel\n  layout are the most common character-material leak — the flat gray studio renders as the\n  actual set. Pair every character-sheet role with its own exclusion: *\"Do not take the\n  gray backdrop, the panel borders, or the multi-view layout.\"*\n- **Beat lines name characters, never handles.** In action\u002Fbeat prose, a character appears\n  as name + one visible marker at their first appearance in the beat — *\"Mira — silver\n  streak, rust-red jacket — crosses the stall line\"* — not as `@Image 2`. The model binds\n  by what it can see in the material, and a handle used as a sentence subject is the\n  classic way one character comes back as two people. `[OFFICIAL — SD25-PE mapping\n  priority: material content outranks upload order]`\n\n### Fidelity — say how much of each material must survive\n\n`[EMPIRICAL — MiniMax H3 skill corpus, re-derived; cross-model structure, unmeasured on\nSeedance]` A role says what job a material does; it still doesn't say how much of the\nmaterial must reach the pixels. Declare one fidelity grade per material, in the same\nsentence as its role:\n\n- **full-preserve** — the subject appears as-is: face, build, wardrobe, all of it.\n- **partial-preserve** — the named parts survive, the rest is free: *\"the jacket and the\n  scar; hairstyle may change.\"*\n- **attribute-transfer** — named traits lift onto a **different, named target**: *\"apply\n  this fabric's weave and sheen to Mira's coat.\"* The target must be named — this is the\n  one case a bare role line cannot express, and the one that goes wrong silently.\n- **loose-guide** — mood, palette, or energy only; nothing is copied literally.\n\n```\n@Image 4 defines the brocade fabric — attribute-transfer onto Mira's coat only.\nDo not carry the garment's cut, the mannequin, or the studio backdrop.\n```\n\n### Material budget\n\n`[OFFICIAL — Dreamina]` Hard limits vs the ranges that actually stay stable:\n\n| Type | Hard limit | Stable range |\n|---|---|---|\n| Images | 30, each ≤4K | 1–8 distinct subjects |\n| Videos | 10, ≤30s combined | 1–5 subjects, 5–10s each |\n| Audio | 10 clips, ≤30s combined | only clips directly relevant |\n| Video-edit source | 1 video + reference images | source ≤20s, 1–5 reference images |\n\n50 reference materials total. Above the stable ranges (9–12 subjects in images, 6–10 in\naudio\u002Fvideo, 6–8 edit reference images) generation still works but stability drops and the\nshot may need several attempts — budget for it, or split the scene.\n\n**More than five subjects needing multiple views → one view per image.** Independent view\nimages beat a single collage of views; the collage is the less stable form.\n\n---\n\n## Multi-Reference — the Five-Step Workflow\n\n`[OFFICIAL — Dreamina]` The goal is **not** to cram every reference into one sentence. It\nis to define the relationships among characters, props, scenes, actions, and audio so the\nmodel picks the right material for the right moment.\n\n**Step 1 — name and map each subject individually.** One line per subject:\n\n```\n\u003CCharacter A> corresponds to @Image 1. Use only the appearance, hairstyle, and clothing.\n\u003CCharacter B> corresponds to @Image 2. Use only the appearance, hairstyle, and clothing.\n\u003CProp A> corresponds to @Image 3. Use only the structure, material, and color.\n\u003CScene A> references @Image 4. Use only the spatial layout, architecture, and lighting.\nDo not use the people in the image.\n```\n\n> **The canonical failure:** `@Images 1 through 4 define four characters respectively.`\n> That sentence does not say which image is which character, and the model will guess.\n\n**Step 2 — group by type** once several subjects are in play: `[Characters]` → `[Props]` →\n`[Scenes]` → `[Motion and Audio]`. Add the non-interchange lock to the character group:\n\"Do not interchange these characters' appearances, clothing, actions, positions, or dialogue.\"\n\n**Step 3 — profile any recurring subject.** A character crossing several scenes, or\ncarrying several materials, gets one consolidated block:\n\n```\n[Subject Profile: Conservator]\nAppearance and clothing: @Image 1.\nFixed prop: \u003CSample Case> from @Image 5.\nLocations: \u003CConservation Lab> and \u003CGallery>.\nMotion references: the case-opening motion from @Video 1.\nDo not use: other characters' clothing. Do not give this character other equipment.\n```\n\n**Step 4 — select references by scene.** Per scene, list only the subset actually used,\nthen the event and its end state:\n\n```\nScene 1 | Inspection in the Conservation Lab\nUse: \u003CConservator>, \u003CSample Case>, \u003CConservation Lab>, and the case-opening motion from @Video 1.\nEvent: \u003CConservator> opens \u003CSample Case> at the workbench and inspects the sample inside.\nEnd state: \u003CConservator> remains on the inner side of the workbench. \u003CSample Case> stays\nbeside the conservator's right hand.\n```\n\n**Step 5 — check ownership.** Props belong to exactly one character (\"belongs only to\n\u003CConservator>\"); character count, clothing, and spatial direction stay constant across scenes.\n\n---\n\n## Long Video — Stages and End States\n\n`[OFFICIAL — Dreamina]` 2.5 generates up to 30 seconds natively. Anything with several\nevents gets **staged** — one flat paragraph is where dropped beats come from.\n\nEach stage carries exactly **one primary state change** and closes on an **explicit,\ndirectly visible end state**. Each new stage restates what carries over.\n\n```\n[Generation Goal]\nGenerate a \u003Cvideo type>. The central subject is \u003Csubject>, and the primary event is \u003Cstory summary>.\n\n[Stage 1]\nInitial state: \u003Cinitial state of characters, props, and scene>.\nPrimary event: \u003Cone primary action or event>.\nEnd state: \u003Ccharacter positions, prop ownership, or visible scene state>.\n\n[Stage 2]\nContinue from the previous stage: \u003Cstate that must remain unchanged>.\nPrimary event: \u003Cone primary action or event>.\nEnd state: \u003Cobservable state>.\n\n[Stage 3]\nPrimary event: \u003Cclosing event>.\nEnd state: \u003Cfinal visible state>.\n\n[Maintain Consistency]\nKeep \u003Ccharacter identity, number of characters, clothing, prop ownership, spatial direction,\nand audio relationships> consistent.\n```\n\n### Timestamps and pacing\n\nStages are the default. Reach for one-second precision only when a specific handoff,\nentrance\u002Fexit, transition, or beat must land on a moment.\n\n| Pattern | Use for | Example |\n|---|---|---|\n| Time range | Allocating pacing to a segment | `0-3 seconds… 3-7 seconds… 7-12 seconds…` |\n| Exact time point | One key event | `At 5 seconds, the camera whip-pans rapidly to the left and completes the transition.` |\n| Relative timing | A delay between two events | `Three seconds after the character presses the button, the room lights gradually turn off.` |\n\n- Ranges are **consecutive and non-overlapping**, and they are a **time budget**, not a\n  frame-accurate edit point — actions may land slightly either side of a boundary.\n- Too little content in a range hands the model freedom you did not intend; too much causes\n  excessive cutting or dropped events.\n- **Never demand a frequency inside one second** (\"complete three actions in one second\").\n  Timestamps allocate time; they do not enforce cadence.\n- Every timestamped segment still ends on an explicit end state, exactly like a stage.\n\n> **Relationship to 2.0's beat arithmetic.** `..\u002Fhiggsfield-seedance\u002FSKILL.md` § Output\n> Format requires timed beats to sum to the declared duration. That still holds — but sum\n> to 4–30s here, and remember that on 2.5 the sum is a budget the model approximates, not a\n> cut list it honours to the frame.\n\n---\n\n## Audio and Text — Bracket Syntax\n\n`[OFFICIAL — Dreamina]` Prompts can be written entirely in natural language. Use the\nbrackets when music, SFX, dialogue, and subtitles must be told apart explicitly:\n\n| Content | Syntax | Example |\n|---|---|---|\n| Music | `()` | `(Soft, rhythmic piano music plays in the background)` |\n| Sound effects | `\u003C>` | `\u003CA bell rings in the distance>` |\n| Dialogue | `{}` | `{Hello, welcome back.}` |\n| Subtitles | `【】` | `【Chapter One: Departure】` |\n\n**Dialogue language reinforcement.** For non-Chinese dialogue — or when English text comes\nback spoken in Chinese, or a specific regional variety is wanted — state it before the line:\n\n```\nDialogue language + regional variety or accent + delivery style + speaker + {dialogue}\n```\n\n```\nDialogue language: authentic Los Angeles English. The young man says in natural Los Angeles\nvernacular: {No way, you actually made it.}\n```\n\nTwo house rules carry over from `..\u002Fhiggsfield-audio\u002FSKILL.md` and the film pipeline in\n`..\u002Fhiggsfield-seedance\u002FHELL-GRIND.md`:\n\n- **Speech lives in the audio clause only** — not a word of dialogue inside the action\n  description, or the model narrates it as behavior. The trap is that writers do not think\n  of *subtext* as speech: `they exchange a look that says \"you too?\"` is a line the script\n  never wrote, and it comes back spoken. `[HOUSE — the failing take is timecoded in the\n  nutllwhy\u002Fseedance-tvc-director evaluation, MIT, 2026-08-09.]` Anything readable is a\n  voicing request — quoted subtext, a remembered line, a slogan, a sign read aloud. Write\n  the visible behavior instead (*jaw sets, eyes hold*), and note that adding \"no dialogue\"\n  does not undo it: the readable text is still there being asked for.\n- **Diegetic-only is a project choice, and 2.5 finally obeys it.** Random subtitles and\n  unrequested BGM were 2.0's most-reported nuisance; ByteDance calls suppression of both a\n  headline 2.5 fix `[OFFICIAL — Dreamina]`. Still say it — `(no music)` and no `【】` block —\n  rather than trusting the improvement.\n\n---\n\n## First-Last Frame and Multi-Keyframe Control\n\n`[OFFICIAL — Dreamina]` On 2.5 these are **prompt statements, not a mode**, because the\nplatform surface has no start\u002Fend-frame media role.\n\n```\n@Image 1 is the first frame. It defines the opening composition, subject position, pose,\nprop state, scene, and camera direction.\n@Image 2 is the last frame. It defines the ending composition, subject position, pose,\nprop state, scene, and camera direction.\n@Image 3 defines \u003CSubject A>'s \u003Cappearance, clothing, structure, or material>. Do not change\nthe first-frame composition defined by @Image 1 or the last-frame composition defined by @Image 2.\n\n\u003CDescribe one continuous action or event>.\nThe video begins naturally from the first frame defined by @Image 1 and reaches the last\nframe defined by @Image 2 after the continuous action.\nBetween the first and last frames, maintain continuity in \u003Ccharacter identity, prop structure\nand ownership, scene layout, and camera direction>.\n```\n\nThree failure sources, all avoidable:\n\n1. **Never merge the anchors** — `@Images 1 and 2 are the first and last frames` is the\n   documented wrong form. One role sentence per image.\n2. **First and last images must share an aspect ratio**, or the last frame stretches. The\n   output ratio locks to the **first** image; duration stays settable.\n3. **Supplementary references supplement only their named attribute** — each one repeats the\n   \"do not change the first\u002Flast-frame composition\" clause.\n\n**Multi-keyframe sequences** (3+ ordered stage images) open with `Use @Image 1 through\n@Image N as keyframes in this order`, then describe the key state each image represents.\nIndependent keyframe images align far more reliably than several frames combined into one\ngrid. Keyframes control **stage order and key states** — they do not reproduce every\nintermediate frame.\n\nStoryboard grids and blockout references (coarse vs fine) are the next rung up:\n`MODE-PLAYBOOKS.md` § Storyboard grids and § Blockout references.\n\n---\n\n## Editing, Extension, and the Composite Modes\n\nThe long templates live in **`MODE-PLAYBOOKS.md`** in this directory. What you need to know\nbefore opening it:\n\n- **Video editing** — define the source as the **sole editing master**, then edit goal, edit\n  scope, target material, and the preserve list. Subject replacement additionally needs a\n  **Timeline Inheritance** clause so the new object inherits the old one's exact timing,\n  path, occlusion, and exit. Background replacement scopes the edit to *outside the\n  subject's silhouette*. Audio categories (dialogue, language, voice, BGM, SFX) edit\n  independently.\n- **Video extension** — the one rule that decides success: **align the boundary frame before\n  describing new content.** Forward extension continues from the source's last frame;\n  backward extension must land *on* the source's first frame as its explicit end state.\n  \"Then connect to the source video\" is the documented failure phrasing — it leaks later\n  elements in early and lets the image drift after it arrives.\n- **One-click video** (images → one edited video) — `Material Roles → Image Order → Motion\n  Amount → Editing Style → Visual Treatment → Audio`. \"Turn these into a video\" is never enough.\n- **Seamless transitions** (bridge content between two clips) — `Before Video → After Video →\n  Trigger Action → Camera Movement → Visual Transformation → Arrival State → Audio`.\n- **Transition vocabulary** — natural shot switch · fade in\u002Fout · stacking · white\u002Fblack flash ·\n  erase · mask · similar-object · action · dynamic relay · zoom in\u002Fout · ink-wash. Each\n  named with its own trigger and arrival composition.\n\n---\n\n## Emotional Direction and Camera Terms\n\n`[OFFICIAL — Dreamina]`\n\n**Emotion.** Abstract words (\"tense\", \"warm\", \"oppressive\") set a direction and leave the\nrest to interpretation. Pair them with directly visible or audible cues — eye movement, brow\ntension, mouth movement, breathing, gaze direction, hand movement. **Two to four cues is\nenough for one transition**; listing every facial detail is counterproductive.\n\n```\nThe overall emotion shifts from \u003Cstarting emotion> to \u003Cending emotion>.\nAfter \u003Ctriggering event>, \u003Csubject> first shows \u003Cimmediate observable reaction>.\nThen, \u003Ceyes, brows, mouth, breathing, gaze, or hand movement> gradually \u003Cchanges>.\nFinally, \u003Csubject> expresses \u003Ctarget emotion> through \u003Crestrained or explicit outward behavior>.\n```\n\nUse event-triggered multi-stage form only when the emotion genuinely changes several times.\nFor muscle-level control beyond this, `..\u002Fhiggsfield-facs\u002FSKILL.md` (AU codes); for the\nbehavior-under-pressure layer that makes the cues *mean* something,\n`..\u002Fhiggsfield-acting\u002FSKILL.md`.\n\n**Camera terms.** Basic language works directly — shot size (extreme wide \u002F wide \u002F medium \u002F\nclose-up \u002F extreme close-up), movement (push in, pull out, pan, lateral move, follow, orbit,\ndive, dolly out, tilt up, handheld shake), position (low angle, overhead, first-person).\n\nPopular techniques (one-take, dolly zoom, aerial, FPV, bullet time, handheld, bounce speed\nramp) also work directly — but with several subjects in frame, still say **which** subject\nthe camera follows or orbits, where the move starts, and where it ends.\n\nFor a **niche or ambiguous term**, keep the term and translate it into an observable result:\n\n```\nCinematography term + target subject + visual change + foreground\u002Fbackground relationship + direction or speed\n```\n\n```\nRack focus: shift focus smoothly from the leaves in the foreground to the person in the\nbackground. The leaves gradually blur while the person's face changes from soft to sharp.\n```\n\nFor a precise transition moment, add the trigger time, the occluding object, the direction,\nand what continues afterwards. Aperture\u002Ffocal-length\u002Fshutter numbers are allowed but the\nintended visible result is clearer than a raw value alone.\n\n> **FOV degrees still beat millimetres.** The discrete-anchor FOV bank in\n> `..\u002Fhiggsfield-seedance\u002FSKILL.md` § FOV anchors is house doctrine across the Seedance\n> family and applies here unchanged.\n\n---\n\n## The Real-Person Character Formula\n\n`[OFFICIAL — Dreamina]` The vendor's answer to \"my characters look AI, or look like twins\"\nis a seven-slot character block. It works for stylized characters too.\n\n```\n[Role]  [Skin color \u002F skin texture]  [Facial details]  [Eyes \u002F soul]\n[Hairstyle \u002F hair color]  [Clothing \u002F clothing texture]  [Body type \u002F mood \u002F temperament]\n[Other requirements, if any]\n```\n\nSlot notes:\n\n- **Skin** — carry real texture: visible micro-pores, translucency, capillary flush. This is\n  the single slot that most separates \"photo of a person\" from \"AI render\".\n- **Facial details** — eye shape, brow line, nose bridge, lips, jawline. Specific beats pretty.\n- **Eyes \u002F soul** — what the gaze is *doing* and what it carries. Dead eyes are the number-one\n  tell; see `..\u002Fhiggsfield-acting\u002FSKILL.md` § Eye life.\n- **Body type \u002F mood \u002F temperament** — the body as biography, not a compliment.\n\n> **House override on slot 1.** The source guide labels the first slot `[Age\u002FRace]`. **Do\n> not write age.** Engine rule 1 in `..\u002Fhiggsfield-seedance\u002FENGINE-RULES.md` is age-blind\n> characters, and Higgsfield's own feature-film brief gives the reason: the content filter\n> tightens sharply the moment it reads a minor. Write **role, build, clothing, and action**\n> instead — \"a lean courier in a soaked parka\", not \"a 22-year-old\". Ethnicity stays a normal\n> descriptive slot; age does not.\n\n---\n\n## Choosing 2.0 vs 2.5\n\n| The job needs | Model |\n|---|---|\n| 4K or 1080p output | **2.0** (`mode=std`) — 2.5 has no lane above 720p |\n| Platform-pinned start \u002F end frame media role | **2.0** |\n| A `genre` hint parameter | **2.0** |\n| A clip longer than 15 seconds in one generation | **2.5** (up to 30s) |\n| Editing a video that already exists | **2.5** (`video_edit`) |\n| Extending a clip forward *or backward* | **2.5** (`video_extension`) |\n| More than 9 reference images, or >3 reference videos\u002Faudio clips | **2.5** (30 \u002F 10 \u002F 10 vs 2.0's 9 \u002F 3 \u002F 3) |\n| Cheap 480p prompt validation | either — both cap the draft lane at 480p |\n\n**The honest default:** draft and structure on 2.5 when the job is long, reference-heavy, or\nedit-shaped; finish on 2.0 when the deliverable needs resolution. They are not\ninterchangeable takes — a 2.5 draft validates the *prompt*, not the 2.0 render, for the same\nreason `..\u002Fhiggsfield-seedance\u002FSKILL.md` § Drafts Validate the Prompt gives: there is no\nseed to pin.\n\n### What carries over from 2.0 unchanged\n\nThe engine rules are shared. `..\u002Fhiggsfield-seedance\u002FENGINE-RULES.md` applies in full —\nage-blind characters, exit-frame = implicit cut, off-screen = nonexistent, avoid reflection\nshots, ≤3 tracked characters, double-contrast cuts, micro-expressions as physics. So do the\npositive-phrasing law (Seedance has no negative-embedding architecture in either version),\nthe homograph trap, and the block scaffold for production-scale briefs.\n\n---\n\n## Hard Limits — Do Not Over-Promise These\n\n`[OFFICIAL — Dreamina]`\n\n- **Timestamps allocate time; they are not frame-accurate edit points.** Never imply a beat\n  will land on an exact frame.\n- **Video-editing prompts raise the probability** that critical events align with the source —\n  they cannot guarantee frame-by-frame overlap.\n- **Multi-reference selects and combines** the right materials per scene; it does not make\n  every material appear at once.\n- **Frame-level text accuracy is not promised.** Subtitles, formulas, signs, and product\n  specs need prepared reference materials plus post — the same conclusion as 2.0's text-rendering\n  high-risk row.\n- **Video editing locks the input's aspect ratio and approximate duration**; output may differ\n  by up to ~0.3s from transition-frame handling.\n- **First\u002Flast-frame generation locks aspect ratio to the first image**; mismatched ratios\n  stretch the last frame.\n- **Video extension locks aspect ratio to the source**; the extended segment's audio volume may\n  differ slightly.\n- **Boundary frames connect naturally, not identically** — review both sides of the seam plus\n  the whole extended segment.\n- **Seamless transitions aim for visual and audio continuity**, not pixel-identical preservation\n  of either source clip.\n\n---\n\n## Dreamina-Only — What Higgsfield Does Not Expose\n\n`[DREAMINA-ONLY]` These exist in ByteDance's own Dreamina\u002FJimeng product and appear in the\nguides, but there is **no Higgsfield parameter behind them** on the 2026-08-07 snapshot. Do\nnot offer them here.\n\n| Dreamina feature | Status on Higgsfield | Closest thing that does work |\n|---|---|---|\n| **Ultra Long Video** — 30–180s in one generation | Not exposed; `duration` caps at 30 | Stage a 30s generation, then chain `video_extension` |\n| **Nested extension to 60s** via repeated UI operations | The *model* rule (source ≤30s → extend up to 30s) holds; the one-click nesting UI does not exist | Chain `video_extension` calls, re-checking the boundary each time |\n| **Edit with marks \u002F Advanced Edit** — box, arrow, brush, anchor annotations on a frame | Not exposed; there is no annotation channel | `video_edit` with a written scope: object + change + effective time range |\n| **Clay Renderer plugin** — white-model rendering workflow | Not exposed as a plugin | Coarse\u002Ffine blockout prompting via `omni_reference` (`MODE-PLAYBOOKS.md` § Blockout references) |\n| Native 180s \u002F multi-minute deliverables | — | Generate in 30s pieces and assemble in post (`..\u002Fhiggsfield-pipeline\u002FSKILL.md`) |\n\n**Extension chain math.** The model-side rule is: a source **within 30 seconds** can be\nextended by 4–30 seconds in one operation, so a 30s source plus a 30s extension is the\n60-second ceiling for a single chain. Beyond that, re-anchor from the original references\nrather than extending an extension — the same degradation curve as 2.0's chain cap in\n`..\u002Fhiggsfield-seedance\u002FSKILL.md` § Extension Prompting.\n\n---\n\n## Pre-Submission Checklist\n\n`[OFFICIAL — Dreamina]`, plus the house preflight. Only the rows for techniques actually used:\n\n- [ ] Subject and primary action or event stated plainly\n- [ ] Every reference states what to use **and** what not to use\n- [ ] Every character, product, and prop named and bound to a specific material\n- [ ] References selected **per scene**, not all required at once\n- [ ] Each long-video stage has one primary change and a clear end state\n- [ ] Character count, clothing, prop ownership, and spatial relationships consistent throughout\n- [ ] Editing prompts define the sole master, edit scope, target quantity, and preserve list\n- [ ] Abstract emotions and niche camera terms paired with observable cues\n- [ ] First\u002Flast frames: one role per image, matching aspect ratios, anchors not merged\n- [ ] Storyboards state which structure to inherit, not literal panel reproduction\n- [ ] Blockouts: coarse vs fine identified first, inheritance list stated\n- [ ] Auto-locked parameters respected for edit \u002F first-last \u002F extension\n- [ ] Extension: boundary frame, motion trend, and audio continuity all checked\n- [ ] One-click video: roles, order, motion amount, editing style, audio all defined\n- [ ] Transitions: both clips' roles, trigger, process, and arrival state defined\n- [ ] No age words anywhere (engine rule 1)\n- [ ] `seedance_lint.py --preflight --model seedance_2_5` run and clean\n\n---\n\n## Related Skills\n\n- `..\u002Fhiggsfield-seedance\u002FSKILL.md` — Seedance 2.0 director: the filter model, block scaffold,\n  FOV anchors, prompt-craft laws, engine rules. **Everything model-agnostic lives there.**\n- `..\u002Fhiggsfield-seedance\u002FENGINE-RULES.md` — the shared hard rendering constraints\n- `..\u002Fhiggsfield-seedance\u002FHELL-GRIND.md` — Higgsfield's open-sourced feature-film pipeline\n  (assets, GEO spatial layout, dialogue construction, iteration discipline)\n- `..\u002Fhiggsfield-acting\u002FSKILL.md` — performance: objective, obstacle, tactics, beats, eye life\n- `..\u002Fhiggsfield-facs\u002FSKILL.md` — facial control by Action Unit code\n- `..\u002Fhiggsfield-seedance-vfx\u002FSKILL.md` — video-to-video footage transforms on 2.0\n- `..\u002Fhiggsfield-models\u002FSKILL.md` — model choice against the specs layer\n- `..\u002Fhiggsfield-pipeline\u002FSKILL.md` — assembling 30-second pieces into a longer deliverable\n- `MODE-PLAYBOOKS.md` — the editing \u002F extension \u002F one-click \u002F transition \u002F blockout templates\n- `VFX-PIPELINE.md` — the AI-VFX production pipeline: asset-class model routing, the size-ref\n  frame, location batching, the `omni_reference` v2v lane, the four-batch rule, the slop catalog\n","Seedance 2.5 is a **different dialect from Seedance 2.0**, not a version bump you can\nprompt through by habit. 2.0 is a reference-driven shot generator with start\u002Fend frames\nand a 4K lane. 2.5 is an **omni-reference production model**: up to 50 reference\nmaterials, 30-second native runtime, and three non-generation modes — it can edit a video\nyou already have, and extend one forward or backward from its boundary frame.\n\nThe prompt grammar changes with it. Reference roles are declared in prose (`@Image 1\ndefines …`), audio and text get bracket syntax, long videos are staged with explicit end\nstates, and first\u002Flast frames are announced **inside the prompt** rather than selected as\na mode.\n\n> **Model split — read this before writing anything.** 2.5 caps at **720p** and has no\n> `start_image` \u002F `end_image` media role, no `genre` hint, and no 4K lane. If the job\n> needs 4K, a genre hint, or platform-level start\u002Fend frame pinning, it is a\n> **Seedance 2.0** job — `..\u002Fhiggsfield-seedance\u002FSKILL.md`. See § Choosing 2.0 vs 2.5.\n\n## QUICK FACTS\n*Generated-checked block (scripts\u002Fbuild_index.py verifies anchors). Routing aids — read the linked sections for the rules themselves.*\n- Four modes, picked **before** writing: `t2v` · `omni_reference` · `video_edit` · `video_extension`; the mode changes what the prompt *is* [→](#the-mode-router)\n- Higgsfield surface: **480p\u002F720p only**, duration **4–30s**, no start\u002Fend-frame role, no genre hint, `extension_mode` required for (and only for) `video_extension` [→](#the-higgsfield-parameter-surface)\n- `video_edit` **ignores** `duration` and `aspect_ratio` and bills by the source video's length; `video_extension` inherits the source's aspect ratio [→](#the-higgsfield-parameter-surface)\n- Every reference material gets an explicit role **and** an exclusion — \"what to use\" plus \"what not to use\"; never let the model infer the mapping [→](#reference-roles-say-what-to-use-and-what-not-to-use)\n- Each material also declares a **fidelity grade** — full-preserve \u002F partial-preserve \u002F attribute-transfer (name the target) \u002F loose-guide; beat lines name characters (name + one visible marker), never handles [→](#fidelity-say-how-much-of-each-material-must-survive)\n- Material budget: 30 images \u002F 10 videos ≤30s total \u002F 10 audio ≤30s total, 50 materials max; stability ranges are 1–8 subjects (images), 1–5 subjects at 5–10s (video\u002Faudio) [→](#material-budget)\n- Multi-reference is a 5-step workflow — map → group → profile → select-by-scene, one line per subject; `@Images 1 through 4 define four characters` is the canonical failure [→](#multi-reference-the-five-step-workflow)\n- Long videos are **staged**, not paragraphed: one primary change per stage + an explicit **end state**; timestamps allocate a budget, they are not frame-accurate edit points [→](#long-video-stages-and-end-states)\n- Bracket syntax: `()` music · `\u003C>` SFX · `{}` dialogue · `【】` subtitles; non-Chinese dialogue needs a language line before the line [→](#audio-and-text-bracket-syntax)\n- First\u002Flast frames and multi-keyframes are declared **in the prompt** (`@Image 1 is the first frame`) — aspect ratio locks to the first image; never merge the two anchors into one sentence [→](#first-last-frame-and-multi-keyframe-control)\n- Editing needs a **sole editing master** + edit scope + Timeline Inheritance; extension needs the **boundary frame aligned before** any new content: `MODE-PLAYBOOKS.md`\n- Storyboard grids, coarse-vs-fine blockouts, one-click video, seamless transitions: `MODE-PLAYBOOKS.md`\n- **AI-VFX production pipeline** — model-per-asset-class routing, the size-ref frame, location batching, the `omni_reference` v2v lane (source ≥4s, duration = source), the four-batch rule, the slop catalog: `VFX-PIPELINE.md`\n- Emotion needs 2–4 **observable** cues, not adjectives; niche camera terms get translated into a visible result [→](#emotional-direction-and-camera-terms)\n- The real-person formula is 7 slots — and slot 1 is **role, never age**: the age-blind engine rule outranks the source guide's `[Age\u002FRace]` label [→](#the-real-person-character-formula)\n- Hard limits that must not be over-promised (frame accuracy, locked parameters, pixel-identical transitions) [→](#hard-limits-do-not-over-promise-these)\n- Dreamina-product features that are **not** on the Higgsfield surface — Ultra Long Video 180s, mark-based editing, Clay Renderer [→](#dreamina-only-what-higgsfield-does-not-expose)\n\n---\n\n## Provenance\n\nTwo independent sources, labelled throughout:\n\n| Label | Source |\n|---|---|\n| `[OFFICIAL — Dreamina]` | ByteDance's *Dreamina Seedance 2.5 Prompt Guide* + *User Guide* — the model vendor's own prompt doctrine. Prompt grammar is model-side, so it carries across to Higgsfield's hosting. |\n| `[OFFICIAL — platform]` | Higgsfield's live `models_explore` catalog, snapshot **2026-08-07** (`..\u002F..\u002Fspecs\u002Fmodel-specs.json`). Parameters, enums, and media roles come from here and nowhere else. |\n| `[DREAMINA-ONLY]` | A Dreamina *product* feature with no Higgsfield parameter behind it. Never quote these as things the user can do here. |\n\nWhere the two disagree about what is *settable*, the platform snapshot wins — it is what\nthe API actually accepts.\n\n---\n\n## The Mode Router\n\nPick the mode first. The same sentence means different things in different modes, and two\nof the four modes are not generation at all.\n\n| The user wants | Mode | What the prompt is |\n|---|---|---|\n| A clip from a description, no materials | `t2v` | A scene brief — the core formula below |\n| A clip built from images \u002F videos \u002F audio they supply | `omni_reference` | A **role map** plus a scene brief |\n| To change something inside a video they already have | `video_edit` | An **edit order**: master + scope + preserve list |\n| More footage before or after a video they already have | `video_extension` | A **boundary contract** plus new content |\n\nTwo rules that fall out of this:\n\n1. **First\u002Flast frames, keyframes, storyboard grids, and blockouts are all `omni_reference`.**\n   2.5 has no separate first\u002Flast-frame mode on this platform — the anchor images are\n   ordinary references whose *role sentence* says they are the first and last frame.\n   `[OFFICIAL — Dreamina: \"no need to switch to a separate first\u002Flast-frame mode\"]`\n2. **Editing is not regeneration.** If the user wants the shot rebuilt, that is\n   `omni_reference` with the old clip as a motion reference — not `video_edit`. `video_edit`\n   preserves the master's timeline and changes one scoped thing inside it.\n\n> **Video-to-video is not automatically `video_edit`.** The field VFX workflow — swap the\n> person in this plate, keep every other pixel — runs in **`omni_reference` with the source\n> attached as a video reference**, because that is the lane where `duration` is settable and\n> must be **matched to the source** (and where the source therefore has to be **≥ 4 s**, the\n> `duration` floor). `video_edit` is the lane for a scoped change inside a master whose\n> timeline must survive untouched. Full routing table + the performance-inheritance clause:\n> `VFX-PIPELINE.md` § Stage 4. `[FIELD — AI-vs-VFX, 2026-08-08]`\n\n---\n\n## The Higgsfield Parameter Surface\n\n`[OFFICIAL — platform, snapshot 2026-08-07]` · verify against `..\u002F..\u002Fspecs\u002Fmodel-specs.json`\nbefore quoting (HARD RULE 3).\n\n| Parameter | Values | Notes |\n|---|---|---|\n| `mode` | `t2v` · `omni_reference` · `video_edit` · `video_extension` | default `t2v` |\n| `duration` | 4–30 s | default 5 — **ignored in `video_edit`** |\n| `resolution` | `480p` · `720p` | default 720p — **there is no 1080p or 4K on 2.5** |\n| `generate_audio` | bool | default true |\n| `extension_mode` | `backward` · `forward` | **required** for `video_extension`, **not allowed** otherwise |\n| aspect ratio | `auto` · `21:9` · `16:9` · `4:3` · `1:1` · `3:4` · `9:16` | ignored in `video_edit`; follows the source in `video_extension` |\n| media roles | `image_references` · `video_references` · `audio_references` | **no `start_image` \u002F `end_image`** |\n\nThree consequences worth stating to the user before they spend credits:\n\n- **`video_edit` bills by the source video's duration**, and neither `duration` nor\n  `aspect_ratio` is settable — a 20-second master costs a 20-second render no matter how\n  small the edit.\n- **`video_extension` inherits the source's aspect ratio**; only the extension's\n  *duration* is yours to set.\n- **No `genre` parameter.** 2.0's genre hint does not exist here — genre lives in the\n  prompt's visual-style clause instead.\n\nPreflight the same way as 2.0:\n\n```\npython3 scripts\u002Fseedance_lint.py --preflight --model seedance_2_5 \"\u003Cprompt>\"\n```\n\nThe linter reads the enums out of `..\u002F..\u002Fspecs\u002Fmodel-specs.json`, so an out-of-range duration, a\n1080p request, or a `video_extension` missing its `extension_mode` is caught before the\nrender.\n\n---\n\n## The Core Prompt Formula\n\n`[OFFICIAL — Dreamina]` Combine only the parts the shot needs; omit the rest rather than\npadding empty slots.\n\n```\n\u003CSubject> performs \u003Cprimary action or event> in \u003Cscene and environment>.\nThe visuals feature \u003Cvisual style>.\nUse \u003Cshot size, camera angle, camera movement, or cuts>.\nAudio includes \u003Cdialogue, ambience, sound effects, or music>.\n```\n\n- **Subject + action** is load-bearing — make it concrete. \"The man runs\" → \"the man\n  accelerates into a sprint while his jacket reacts to the airflow.\"\n- **Scene and environment** — location, time, weather, spatial relationships, background state.\n- **Visual style** — lighting, color, materials, texture, mood. Only descriptors that add\n  information; stacked buzzwords (\"cinematic, 8K, masterpiece\") sample nothing in particular.\n  The named-substitute discipline in `..\u002Fhiggsfield-seedance\u002FSKILL.md` § Prompt-Craft Laws\n  applies unchanged.\n- **Camera** — shot size, angle, movement, focus subject, transitions. Motion matches the\n  action; it is not decoration.\n- **Audio** — dialogue, voice characteristics, ambience, SFX, music, synchronized to the visuals.\n\n**Generation parameters are not prompt text.** Resolution, duration, and aspect ratio are\nset on the generation page or via the API — writing them into the prose does nothing except\nin the modes that auto-lock them, where they are not settable at all.\n\n---\n\n## Reference Roles — Say What to Use *and* What Not to Use\n\nThe moment there is more than one material, or a material sitting next to a text\ndescription, the prompt must state what each material contributes. `[OFFICIAL — Dreamina]`\n\n```\n@Image 1 defines \u003Csubject>'s \u003Cappearance, clothing, structure, or material>.\n@Video 1 defines \u003Cmotion, camera movement, or pacing>.\n@Audio 1 defines \u003Ccharacter or sound type>'s \u003Cvoice, dialogue, ambience, or music>.\n```\n\nEvery material that could leak something unwanted gets an explicit exclusion in the same\nsentence:\n\n```\n@Image 2 defines the workbench and window light. Do not use the people in the image.\n```\n\nRules:\n\n- **Mappings live in the prompt.** Text labels drawn inside an image are not a mapping, and\n  the model will not infer which person or prop a material represents.\n- **Video-only references are motion\u002Fpacing references by default**, not identity — say so\n  explicitly when you mean otherwise.\n- **Several views of one subject must say they are one subject**: \"All four images define\n  one folding desk lamp. The output must contain only one lamp throughout.\" Without it the\n  model duplicates the subject.\n- **When a reference video already carries the motion accurately, state only which\n  attributes to inherit.** Restating every action fights the reference. A blockout or motion\n  video carries motion and spatial structure — not identity — so the prompt still has to\n  define subjects, scene, action, and visual style.\n- **Never place a reference handle in a shot where that subject is absent** — the same rule\n  as 2.0's tag discipline; the model forces it into frame.\n- **Character sheets leak their staging.** A sheet's neutral backdrop and multi-view panel\n  layout are the most common character-material leak — the flat gray studio renders as the\n  actual set. Pair every character-sheet role with its own exclusion: *\"Do not take the\n  gray backdrop, the panel borders, or the multi-view layout.\"*\n- **Beat lines name characters, never handles.** In action\u002Fbeat prose, a character appears\n  as name + one visible marker at their first appearance in the beat — *\"Mira — silver\n  streak, rust-red jacket — crosses the stall line\"* — not as `@Image 2`. The model binds\n  by what it can see in the material, and a handle used as a sentence subject is the\n  classic way one character comes back as two people. `[OFFICIAL — SD25-PE mapping\n  priority: material content outranks upload order]`\n\n### Fidelity — say how much of each material must survive\n\n`[EMPIRICAL — MiniMax H3 skill corpus, re-derived; cross-model structure, unmeasured on\nSeedance]` A role says what job a material does; it still doesn't say how much of the\nmaterial must reach the pixels. Declare one fidelity grade per material, in the same\nsentence as its role:\n\n- **full-preserve** — the subject appears as-is: face, build, wardrobe, all of it.\n- **partial-preserve** — the named parts survive, the rest is free: *\"the jacket and the\n  scar; hairstyle may change.\"*\n- **attribute-transfer** — named traits lift onto a **different, named target**: *\"apply\n  this fabric's weave and sheen to Mira's coat.\"* The target must be named — this is the\n  one case a bare role line cannot express, and the one that goes wrong silently.\n- **loose-guide** — mood, palette, or energy only; nothing is copied literally.\n\n```\n@Image 4 defines the brocade fabric — attribute-transfer onto Mira's coat only.\nDo not carry the garment's cut, the mannequin, or the studio backdrop.\n```\n\n### Material budget\n\n`[OFFICIAL — Dreamina]` Hard limits vs the ranges that actually stay stable:\n\n| Type | Hard limit | Stable range |\n|---|---|---|\n| Images | 30, each ≤4K | 1–8 distinct subjects |\n| Videos | 10, ≤30s combined | 1–5 subjects, 5–10s each |\n| Audio | 10 clips, ≤30s combined | only clips directly relevant |\n| Video-edit source | 1 video + reference images | source ≤20s, 1–5 reference images |\n\n50 reference materials total. Above the stable ranges (9–12 subjects in images, 6–10 in\naudio\u002Fvideo, 6–8 edit reference images) generation still works but stability drops and the\nshot may need several attempts — budget for it, or split the scene.\n\n**More than five subjects needing multiple views → one view per image.** Independent view\nimages beat a single collage of views; the collage is the less stable form.\n\n---\n\n## Multi-Reference — the Five-Step Workflow\n\n`[OFFICIAL — Dreamina]` The goal is **not** to cram every reference into one sentence. It\nis to define the relationships among characters, props, scenes, actions, and audio so the\nmodel picks the right material for the right moment.\n\n**Step 1 — name and map each subject individually.** One line per subject:\n\n```\n\u003CCharacter A> corresponds to @Image 1. Use only the appearance, hairstyle, and clothing.\n\u003CCharacter B> corresponds to @Image 2. Use only the appearance, hairstyle, and clothing.\n\u003CProp A> corresponds to @Image 3. Use only the structure, material, and color.\n\u003CScene A> references @Image 4. Use only the spatial layout, architecture, and lighting.\nDo not use the people in the image.\n```\n\n> **The canonical failure:** `@Images 1 through 4 define four characters respectively.`\n> That sentence does not say which image is which character, and the model will guess.\n\n**Step 2 — group by type** once several subjects are in play: `[Characters]` → `[Props]` →\n`[Scenes]` → `[Motion and Audio]`. Add the non-interchange lock to the character group:\n\"Do not interchange these characters' appearances, clothing, actions, positions, or dialogue.\"\n\n**Step 3 — profile any recurring subject.** A character crossing several scenes, or\ncarrying several materials, gets one consolidated block:\n\n```\n[Subject Profile: Conservator]\nAppearance and clothing: @Image 1.\nFixed prop: \u003CSample Case> from @Image 5.\nLocations: \u003CConservation Lab> and \u003CGallery>.\nMotion references: the case-opening motion from @Video 1.\nDo not use: other characters' clothing. Do not give this character other equipment.\n```\n\n**Step 4 — select references by scene.** Per scene, list only the subset actually used,\nthen the event and its end state:\n\n```\nScene 1 | Inspection in the Conservation Lab\nUse: \u003CConservator>, \u003CSample Case>, \u003CConservation Lab>, and the case-opening motion from @Video 1.\nEvent: \u003CConservator> opens \u003CSample Case> at the workbench and inspects the sample inside.\nEnd state: \u003CConservator> remains on the inner side of the workbench. \u003CSample Case> stays\nbeside the conservator's right hand.\n```\n\n**Step 5 — check ownership.** Props belong to exactly one character (\"belongs only to\n\u003CConservator>\"); character count, clothing, and spatial direction stay constant across scenes.\n\n---\n\n## Long Video — Stages and End States\n\n`[OFFICIAL — Dreamina]` 2.5 generates up to 30 seconds natively. Anything with several\nevents gets **staged** — one flat paragraph is where dropped beats come from.\n\nEach stage carries exactly **one primary state change** and closes on an **explicit,\ndirectly visible end state**. Each new stage restates what carries over.\n\n```\n[Generation Goal]\nGenerate a \u003Cvideo type>. The central subject is \u003Csubject>, and the primary event is \u003Cstory summary>.\n\n[Stage 1]\nInitial state: \u003Cinitial state of characters, props, and scene>.\nPrimary event: \u003Cone primary action or event>.\nEnd state: \u003Ccharacter positions, prop ownership, or visible scene state>.\n\n[Stage 2]\nContinue from the previous stage: \u003Cstate that must remain unchanged>.\nPrimary event: \u003Cone primary action or event>.\nEnd state: \u003Cobservable state>.\n\n[Stage 3]\nPrimary event: \u003Cclosing event>.\nEnd state: \u003Cfinal visible state>.\n\n[Maintain Consistency]\nKeep \u003Ccharacter identity, number of characters, clothing, prop ownership, spatial direction,\nand audio relationships> consistent.\n```\n\n### Timestamps and pacing\n\nStages are the default. Reach for one-second precision only when a specific handoff,\nentrance\u002Fexit, transition, or beat must land on a moment.\n\n| Pattern | Use for | Example |\n|---|---|---|\n| Time range | Allocating pacing to a segment | `0-3 seconds… 3-7 seconds… 7-12 seconds…` |\n| Exact time point | One key event | `At 5 seconds, the camera whip-pans rapidly to the left and completes the transition.` |\n| Relative timing | A delay between two events | `Three seconds after the character presses the button, the room lights gradually turn off.` |\n\n- Ranges are **consecutive and non-overlapping**, and they are a **time budget**, not a\n  frame-accurate edit point — actions may land slightly either side of a boundary.\n- Too little content in a range hands the model freedom you did not intend; too much causes\n  excessive cutting or dropped events.\n- **Never demand a frequency inside one second** (\"complete three actions in one second\").\n  Timestamps allocate time; they do not enforce cadence.\n- Every timestamped segment still ends on an explicit end state, exactly like a stage.\n\n> **Relationship to 2.0's beat arithmetic.** `..\u002Fhiggsfield-seedance\u002FSKILL.md` § Output\n> Format requires timed beats to sum to the declared duration. That still holds — but sum\n> to 4–30s here, and remember that on 2.5 the sum is a budget the model approximates, not a\n> cut list it honours to the frame.\n\n---\n\n## Audio and Text — Bracket Syntax\n\n`[OFFICIAL — Dreamina]` Prompts can be written entirely in natural language. Use the\nbrackets when music, SFX, dialogue, and subtitles must be told apart explicitly:\n\n| Content | Syntax | Example |\n|---|---|---|\n| Music | `()` | `(Soft, rhythmic piano music plays in the background)` |\n| Sound effects | `\u003C>` | `\u003CA bell rings in the distance>` |\n| Dialogue | `{}` | `{Hello, welcome back.}` |\n| Subtitles | `【】` | `【Chapter One: Departure】` |\n\n**Dialogue language reinforcement.** For non-Chinese dialogue — or when English text comes\nback spoken in Chinese, or a specific regional variety is wanted — state it before the line:\n\n```\nDialogue language + regional variety or accent + delivery style + speaker + {dialogue}\n```\n\n```\nDialogue language: authentic Los Angeles English. The young man says in natural Los Angeles\nvernacular: {No way, you actually made it.}\n```\n\nTwo house rules carry over from `..\u002Fhiggsfield-audio\u002FSKILL.md` and the film pipeline in\n`..\u002Fhiggsfield-seedance\u002FHELL-GRIND.md`:\n\n- **Speech lives in the audio clause only** — not a word of dialogue inside the action\n  description, or the model narrates it as behavior. The trap is that writers do not think\n  of *subtext* as speech: `they exchange a look that says \"you too?\"` is a line the script\n  never wrote, and it comes back spoken. `[HOUSE — the failing take is timecoded in the\n  nutllwhy\u002Fseedance-tvc-director evaluation, MIT, 2026-08-09.]` Anything readable is a\n  voicing request — quoted subtext, a remembered line, a slogan, a sign read aloud. Write\n  the visible behavior instead (*jaw sets, eyes hold*), and note that adding \"no dialogue\"\n  does not undo it: the readable text is still there being asked for.\n- **Diegetic-only is a project choice, and 2.5 finally obeys it.** Random subtitles and\n  unrequested BGM were 2.0's most-reported nuisance; ByteDance calls suppression of both a\n  headline 2.5 fix `[OFFICIAL — Dreamina]`. Still say it — `(no music)` and no `【】` block —\n  rather than trusting the improvement.\n\n---\n\n## First-Last Frame and Multi-Keyframe Control\n\n`[OFFICIAL — Dreamina]` On 2.5 these are **prompt statements, not a mode**, because the\nplatform surface has no start\u002Fend-frame media role.\n\n```\n@Image 1 is the first frame. It defines the opening composition, subject position, pose,\nprop state, scene, and camera direction.\n@Image 2 is the last frame. It defines the ending composition, subject position, pose,\nprop state, scene, and camera direction.\n@Image 3 defines \u003CSubject A>'s \u003Cappearance, clothing, structure, or material>. Do not change\nthe first-frame composition defined by @Image 1 or the last-frame composition defined by @Image 2.\n\n\u003CDescribe one continuous action or event>.\nThe video begins naturally from the first frame defined by @Image 1 and reaches the last\nframe defined by @Image 2 after the continuous action.\nBetween the first and last frames, maintain continuity in \u003Ccharacter identity, prop structure\nand ownership, scene layout, and camera direction>.\n```\n\nThree failure sources, all avoidable:\n\n1. **Never merge the anchors** — `@Images 1 and 2 are the first and last frames` is the\n   documented wrong form. One role sentence per image.\n2. **First and last images must share an aspect ratio**, or the last frame stretches. The\n   output ratio locks to the **first** image; duration stays settable.\n3. **Supplementary references supplement only their named attribute** — each one repeats the\n   \"do not change the first\u002Flast-frame composition\" clause.\n\n**Multi-keyframe sequences** (3+ ordered stage images) open with `Use @Image 1 through\n@Image N as keyframes in this order`, then describe the key state each image represents.\nIndependent keyframe images align far more reliably than several frames combined into one\ngrid. Keyframes control **stage order and key states** — they do not reproduce every\nintermediate frame.\n\nStoryboard grids and blockout references (coarse vs fine) are the next rung up:\n`MODE-PLAYBOOKS.md` § Storyboard grids and § Blockout references.\n\n---\n\n## Editing, Extension, and the Composite Modes\n\nThe long templates live in **`MODE-PLAYBOOKS.md`** in this directory. What you need to know\nbefore opening it:\n\n- **Video editing** — define the source as the **sole editing master**, then edit goal, edit\n  scope, target material, and the preserve list. Subject replacement additionally needs a\n  **Timeline Inheritance** clause so the new object inherits the old one's exact timing,\n  path, occlusion, and exit. Background replacement scopes the edit to *outside the\n  subject's silhouette*. Audio categories (dialogue, language, voice, BGM, SFX) edit\n  independently.\n- **Video extension** — the one rule that decides success: **align the boundary frame before\n  describing new content.** Forward extension continues from the source's last frame;\n  backward extension must land *on* the source's first frame as its explicit end state.\n  \"Then connect to the source video\" is the documented failure phrasing — it leaks later\n  elements in early and lets the image drift after it arrives.\n- **One-click video** (images → one edited video) — `Material Roles → Image Order → Motion\n  Amount → Editing Style → Visual Treatment → Audio`. \"Turn these into a video\" is never enough.\n- **Seamless transitions** (bridge content between two clips) — `Before Video → After Video →\n  Trigger Action → Camera Movement → Visual Transformation → Arrival State → Audio`.\n- **Transition vocabulary** — natural shot switch · fade in\u002Fout · stacking · white\u002Fblack flash ·\n  erase · mask · similar-object · action · dynamic relay · zoom in\u002Fout · ink-wash. Each\n  named with its own trigger and arrival composition.\n\n---\n\n## Emotional Direction and Camera Terms\n\n`[OFFICIAL — Dreamina]`\n\n**Emotion.** Abstract words (\"tense\", \"warm\", \"oppressive\") set a direction and leave the\nrest to interpretation. Pair them with directly visible or audible cues — eye movement, brow\ntension, mouth movement, breathing, gaze direction, hand movement. **Two to four cues is\nenough for one transition**; listing every facial detail is counterproductive.\n\n```\nThe overall emotion shifts from \u003Cstarting emotion> to \u003Cending emotion>.\nAfter \u003Ctriggering event>, \u003Csubject> first shows \u003Cimmediate observable reaction>.\nThen, \u003Ceyes, brows, mouth, breathing, gaze, or hand movement> gradually \u003Cchanges>.\nFinally, \u003Csubject> expresses \u003Ctarget emotion> through \u003Crestrained or explicit outward behavior>.\n```\n\nUse event-triggered multi-stage form only when the emotion genuinely changes several times.\nFor muscle-level control beyond this, `..\u002Fhiggsfield-facs\u002FSKILL.md` (AU codes); for the\nbehavior-under-pressure layer that makes the cues *mean* something,\n`..\u002Fhiggsfield-acting\u002FSKILL.md`.\n\n**Camera terms.** Basic language works directly — shot size (extreme wide \u002F wide \u002F medium \u002F\nclose-up \u002F extreme close-up), movement (push in, pull out, pan, lateral move, follow, orbit,\ndive, dolly out, tilt up, handheld shake), position (low angle, overhead, first-person).\n\nPopular techniques (one-take, dolly zoom, aerial, FPV, bullet time, handheld, bounce speed\nramp) also work directly — but with several subjects in frame, still say **which** subject\nthe camera follows or orbits, where the move starts, and where it ends.\n\nFor a **niche or ambiguous term**, keep the term and translate it into an observable result:\n\n```\nCinematography term + target subject + visual change + foreground\u002Fbackground relationship + direction or speed\n```\n\n```\nRack focus: shift focus smoothly from the leaves in the foreground to the person in the\nbackground. The leaves gradually blur while the person's face changes from soft to sharp.\n```\n\nFor a precise transition moment, add the trigger time, the occluding object, the direction,\nand what continues afterwards. Aperture\u002Ffocal-length\u002Fshutter numbers are allowed but the\nintended visible result is clearer than a raw value alone.\n\n> **FOV degrees still beat millimetres.** The discrete-anchor FOV bank in\n> `..\u002Fhiggsfield-seedance\u002FSKILL.md` § FOV anchors is house doctrine across the Seedance\n> family and applies here unchanged.\n\n---\n\n## The Real-Person Character Formula\n\n`[OFFICIAL — Dreamina]` The vendor's answer to \"my characters look AI, or look like twins\"\nis a seven-slot character block. It works for stylized characters too.\n\n```\n[Role]  [Skin color \u002F skin texture]  [Facial details]  [Eyes \u002F soul]\n[Hairstyle \u002F hair color]  [Clothing \u002F clothing texture]  [Body type \u002F mood \u002F temperament]\n[Other requirements, if any]\n```\n\nSlot notes:\n\n- **Skin** — carry real texture: visible micro-pores, translucency, capillary flush. This is\n  the single slot that most separates \"photo of a person\" from \"AI render\".\n- **Facial details** — eye shape, brow line, nose bridge, lips, jawline. Specific beats pretty.\n- **Eyes \u002F soul** — what the gaze is *doing* and what it carries. Dead eyes are the number-one\n  tell; see `..\u002Fhiggsfield-acting\u002FSKILL.md` § Eye life.\n- **Body type \u002F mood \u002F temperament** — the body as biography, not a compliment.\n\n> **House override on slot 1.** The source guide labels the first slot `[Age\u002FRace]`. **Do\n> not write age.** Engine rule 1 in `..\u002Fhiggsfield-seedance\u002FENGINE-RULES.md` is age-blind\n> characters, and Higgsfield's own feature-film brief gives the reason: the content filter\n> tightens sharply the moment it reads a minor. Write **role, build, clothing, and action**\n> instead — \"a lean courier in a soaked parka\", not \"a 22-year-old\". Ethnicity stays a normal\n> descriptive slot; age does not.\n\n---\n\n## Choosing 2.0 vs 2.5\n\n| The job needs | Model |\n|---|---|\n| 4K or 1080p output | **2.0** (`mode=std`) — 2.5 has no lane above 720p |\n| Platform-pinned start \u002F end frame media role | **2.0** |\n| A `genre` hint parameter | **2.0** |\n| A clip longer than 15 seconds in one generation | **2.5** (up to 30s) |\n| Editing a video that already exists | **2.5** (`video_edit`) |\n| Extending a clip forward *or backward* | **2.5** (`video_extension`) |\n| More than 9 reference images, or >3 reference videos\u002Faudio clips | **2.5** (30 \u002F 10 \u002F 10 vs 2.0's 9 \u002F 3 \u002F 3) |\n| Cheap 480p prompt validation | either — both cap the draft lane at 480p |\n\n**The honest default:** draft and structure on 2.5 when the job is long, reference-heavy, or\nedit-shaped; finish on 2.0 when the deliverable needs resolution. They are not\ninterchangeable takes — a 2.5 draft validates the *prompt*, not the 2.0 render, for the same\nreason `..\u002Fhiggsfield-seedance\u002FSKILL.md` § Drafts Validate the Prompt gives: there is no\nseed to pin.\n\n### What carries over from 2.0 unchanged\n\nThe engine rules are shared. `..\u002Fhiggsfield-seedance\u002FENGINE-RULES.md` applies in full —\nage-blind characters, exit-frame = implicit cut, off-screen = nonexistent, avoid reflection\nshots, ≤3 tracked characters, double-contrast cuts, micro-expressions as physics. So do the\npositive-phrasing law (Seedance has no negative-embedding architecture in either version),\nthe homograph trap, and the block scaffold for production-scale briefs.\n\n---\n\n## Hard Limits — Do Not Over-Promise These\n\n`[OFFICIAL — Dreamina]`\n\n- **Timestamps allocate time; they are not frame-accurate edit points.** Never imply a beat\n  will land on an exact frame.\n- **Video-editing prompts raise the probability** that critical events align with the source —\n  they cannot guarantee frame-by-frame overlap.\n- **Multi-reference selects and combines** the right materials per scene; it does not make\n  every material appear at once.\n- **Frame-level text accuracy is not promised.** Subtitles, formulas, signs, and product\n  specs need prepared reference materials plus post — the same conclusion as 2.0's text-rendering\n  high-risk row.\n- **Video editing locks the input's aspect ratio and approximate duration**; output may differ\n  by up to ~0.3s from transition-frame handling.\n- **First\u002Flast-frame generation locks aspect ratio to the first image**; mismatched ratios\n  stretch the last frame.\n- **Video extension locks aspect ratio to the source**; the extended segment's audio volume may\n  differ slightly.\n- **Boundary frames connect naturally, not identically** — review both sides of the seam plus\n  the whole extended segment.\n- **Seamless transitions aim for visual and audio continuity**, not pixel-identical preservation\n  of either source clip.\n\n---\n\n## Dreamina-Only — What Higgsfield Does Not Expose\n\n`[DREAMINA-ONLY]` These exist in ByteDance's own Dreamina\u002FJimeng product and appear in the\nguides, but there is **no Higgsfield parameter behind them** on the 2026-08-07 snapshot. Do\nnot offer them here.\n\n| Dreamina feature | Status on Higgsfield | Closest thing that does work |\n|---|---|---|\n| **Ultra Long Video** — 30–180s in one generation | Not exposed; `duration` caps at 30 | Stage a 30s generation, then chain `video_extension` |\n| **Nested extension to 60s** via repeated UI operations | The *model* rule (source ≤30s → extend up to 30s) holds; the one-click nesting UI does not exist | Chain `video_extension` calls, re-checking the boundary each time |\n| **Edit with marks \u002F Advanced Edit** — box, arrow, brush, anchor annotations on a frame | Not exposed; there is no annotation channel | `video_edit` with a written scope: object + change + effective time range |\n| **Clay Renderer plugin** — white-model rendering workflow | Not exposed as a plugin | Coarse\u002Ffine blockout prompting via `omni_reference` (`MODE-PLAYBOOKS.md` § Blockout references) |\n| Native 180s \u002F multi-minute deliverables | — | Generate in 30s pieces and assemble in post (`..\u002Fhiggsfield-pipeline\u002FSKILL.md`) |\n\n**Extension chain math.** The model-side rule is: a source **within 30 seconds** can be\nextended by 4–30 seconds in one operation, so a 30s source plus a 30s extension is the\n60-second ceiling for a single chain. Beyond that, re-anchor from the original references\nrather than extending an extension — the same degradation curve as 2.0's chain cap in\n`..\u002Fhiggsfield-seedance\u002FSKILL.md` § Extension Prompting.\n\n---\n\n## Pre-Submission Checklist\n\n`[OFFICIAL — Dreamina]`, plus the house preflight. Only the rows for techniques actually used:\n\n- [ ] Subject and primary action or event stated plainly\n- [ ] Every reference states what to use **and** what not to use\n- [ ] Every character, product, and prop named and bound to a specific material\n- [ ] References selected **per scene**, not all required at once\n- [ ] Each long-video stage has one primary change and a clear end state\n- [ ] Character count, clothing, prop ownership, and spatial relationships consistent throughout\n- [ ] Editing prompts define the sole master, edit scope, target quantity, and preserve list\n- [ ] Abstract emotions and niche camera terms paired with observable cues\n- [ ] First\u002Flast frames: one role per image, matching aspect ratios, anchors not merged\n- [ ] Storyboards state which structure to inherit, not literal panel reproduction\n- [ ] Blockouts: coarse vs fine identified first, inheritance list stated\n- [ ] Auto-locked parameters respected for edit \u002F first-last \u002F extension\n- [ ] Extension: boundary frame, motion trend, and audio continuity all checked\n- [ ] One-click video: roles, order, motion amount, editing style, audio all defined\n- [ ] Transitions: both clips' roles, trigger, process, and arrival state defined\n- [ ] No age words anywhere (engine rule 1)\n- [ ] `seedance_lint.py --preflight --model seedance_2_5` run and clean\n\n---\n\n## Related Skills\n\n- `..\u002Fhiggsfield-seedance\u002FSKILL.md` — Seedance 2.0 director: the filter model, block scaffold,\n  FOV anchors, prompt-craft laws, engine rules. **Everything model-agnostic lives there.**\n- `..\u002Fhiggsfield-seedance\u002FENGINE-RULES.md` — the shared hard rendering constraints\n- `..\u002Fhiggsfield-seedance\u002FHELL-GRIND.md` — Higgsfield's open-sourced feature-film pipeline\n  (assets, GEO spatial layout, dialogue construction, iteration discipline)\n- `..\u002Fhiggsfield-acting\u002FSKILL.md` — performance: objective, obstacle, tactics, beats, eye life\n- `..\u002Fhiggsfield-facs\u002FSKILL.md` — facial control by Action Unit code\n- `..\u002Fhiggsfield-seedance-vfx\u002FSKILL.md` — video-to-video footage transforms on 2.0\n- `..\u002Fhiggsfield-models\u002FSKILL.md` — model choice against the specs layer\n- `..\u002Fhiggsfield-pipeline\u002FSKILL.md` — assembling 30-second pieces into a longer deliverable\n- `MODE-PLAYBOOKS.md` — the editing \u002F extension \u002F one-click \u002F transition \u002F blockout templates\n- `VFX-PIPELINE.md` — the AI-VFX production pipeline: asset-class model routing, the size-ref\n  frame, location batching, the `omni_reference` v2v lane, the four-batch rule, the slop catalog\n",{},"main","19a54ac988e9929160334271f5c1924ddd9b227e","_modules\u002Fhiggsfield-seedance-2-5",19,"https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill","https:\u002F\u002Fgithub.com\u002Falicewe1","写一些自己用得上的工具，能自动化就绝不手动 · 自动化 \u002F 工程效率 \u002F Agent 技能生态 · 保持好奇，持续构建","2026-09-30 00:41:20",{"source":7,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add alicewe1\u002Falice_skill --skill higgsfield-seedance-2-5",{"tool":37,"command":38},"bunx","bunx skills add alicewe1\u002Falice_skill --skill higgsfield-seedance-2-5",{"tool":40,"command":41},"pnpm","pnpm dlx skills add alicewe1\u002Falice_skill --skill higgsfield-seedance-2-5","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill\u002Fblob\u002Fmain\u002F_modules\u002Fhiggsfield-seedance-2-5\u002FSKILL.md",true,[45,51],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"MODE-PLAYBOOKS.md","REFERENCE","参考文档",23972,"https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill\u002Fblob\u002Fmain\u002F_modules\u002Fhiggsfield-seedance-2-5\u002FMODE-PLAYBOOKS.md",{"path":52,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":53,"sourceHtmlUrl":54},"VFX-PIPELINE.md",27291,"https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill\u002Fblob\u002Fmain\u002F_modules\u002Fhiggsfield-seedance-2-5\u002FVFX-PIPELINE.md",[56,63,70,77,84,91],{"skillNo":57,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":18},"01M7DBMD9P6R27SWQCW5HYVWG0","alice","Alice 六类技能总路由。用户输入 Alice 或要求查看 Alice 技能菜单时，严格输出固定菜单；输入攻\u002F防模式、六类入口或直接描述任务时，按类别→领域→模块渐进路由。","https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill\u002Fblob\u002Fmain\u002Faliceskill\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":66,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":18},"01M7DBM8X5M7YXPYBVFMWF8HRC","alice-reverse","逆向分析渐进路由（二进制 \u002F 移动端 \u002F 协议 \u002F 取证 \u002F 保护机制 \u002F 运行时观测）。触发词：二进制\u002F移动端\u002F反编译\u002FFrida\u002FIDA\u002FGhidra\u002FPE\u002FELF\u002FAPK\u002F协议\u002F取证；先选领域索引，再按任务匹配度与评分选模块，禁止一次读取全部领域。","https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill\u002Fblob\u002Fmain\u002Falice-reverse\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":18},"01M7DBM630E2YKQP29194TT7HC","alice-pentest","web安全渐进路由（Web \u002F API \u002F 资产发现 \u002F 漏洞验证 \u002F 云与网络安全）。触发词：Web\u002FAPI\u002F资产发现\u002F端口\u002F漏洞验证\u002F云与网络安全；先选领域索引，再按任务匹配度与评分选模块，禁止一次读取全部领域。","https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill\u002Fblob\u002Fmain\u002Falice-pentest\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":18},"01M7DBM20M2VDYV5D3MYDGBTQX","alice-game","游戏攻防渐进路由（游戏客户端 \u002F 内存 \u002F 图形 \u002F 反作弊 \u002F 安全研究）。触发词：游戏客户端\u002F内存\u002F图形\u002F偏移\u002F反作弊\u002F模拟器\u002F覆盖层；先选领域索引，再按任务匹配度与评分选模块，禁止一次读取全部领域。","https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill\u002Fblob\u002Fmain\u002Falice-game\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":18},"01M7DBKXVEKB8931MZG23VACFN","alice-crack","卡密授权渐进路由（卡密 \u002F 激活 \u002F 授权验证 \u002F 许可证管理 \u002F 本地授权分析）。触发词：卡密\u002F激活\u002F授权验证\u002Flicense\u002F机器码\u002F试用状态；先选领域索引，再按任务匹配度与评分选模块，禁止一次读取全部领域。","https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill\u002Fblob\u002Fmain\u002Falice-crack\u002FSKILL.md",[],[],{"skillNo":92,"name":93,"nameZh":6,"description":94,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":95,"categories":96,"keywords":97,"publishedAt":18},"01M7DBKTMT90MGV2WP8K73TMQD","alice-assist","技能指令渐进路由（技能路由 \u002F 评分 \u002F 吸收 \u002F 迁移 \u002F 自动化 \u002F 文档与平台集成）。触发词：技能创建\u002F吸收\u002F评分\u002F迁移\u002F自动化\u002F文档\u002F平台指令；先选领域索引，再按任务匹配度与评分选模块，禁止一次读取全部领域。","https:\u002F\u002Fgithub.com\u002Falicewe1\u002Falice_skill\u002Fblob\u002Fmain\u002Falice-assist\u002FSKILL.md",[],[]]