Higgsfield Seedance 2 5

Higgsfield Seedance 2.5 Director

Seedance 2.5 is a different dialect from Seedance 2.0, not a version bump you can prompt through by habit. 2.0 is a reference-driven shot generator with a 4K lane and a genre hint. 2.5 is an omni-reference production model: up to 50 reference materials, 30-second native runtime, and three non-generation modes — it can edit a video you already have, and extend one forward or backward from its boundary frame.

The prompt grammar changes with it. Reference roles are declared in prose (@Image 1 defines …), audio and text get bracket syntax, long videos are staged with explicit end states, and first/last frames live inside omni_reference — either as the platform's start_image / end_image roles or announced inside the prompt — never as a separate mode.

Model split — read this before writing anything. 2.5 tops out at 1080p (no 4K lane) and has no genre hint. Since the 2026-09-26 snapshot it does expose start_image / end_image roles — but only in omni_reference mode. If the job needs 4K or a genre hint, it is a Seedance 2.0 job — ../higgsfield-seedance/SKILL.md. See § Choosing 2.0 vs 2.5.

QUICK FACTS

Generated-checked block (scripts/build_index.py verifies anchors). Routing aids — read the linked sections for the rules themselves.

• Four modes, picked before writing: t2v · omni_reference · video_edit · video_extension; the mode changes what the prompt is →
• Higgsfield surface: 480p/720p/1080p (no 4K), duration 4–30s, start_image/end_image only in omni_reference, t2v takes zero references, no genre hint, extension_mode required for (and only for) video_extension →
• video_edit ignores duration and aspect_ratio and bills by the source video's length; video_extension inherits the source's aspect ratio →
• 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 →
• Each material also declares a fidelity grade — full-preserve / partial-preserve / attribute-transfer (name the target) / loose-guide; beat lines name characters (name + one visible marker), never handles →
• Material budget: 30 images / 10 videos ≤30s total / 10 audio ≤30s total, 50 materials max (Dreamina's figures — the platform enforces only the 30-image and 50-material caps) — a platform start_image/end_image counts against both; stability ranges are 1–8 subjects (images), 1–5 subjects at 5–10s (video/audio) →
• 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 →
• 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 →
• Staging fixes too many EVENTS; two incompatible JOBS in one generation (physics + performance) is a separate cut — split into two prompts and stitch →
• Bracket syntax: () music · <> SFX · {} dialogue · 【】 subtitles; non-Chinese dialogue needs a language line before the line; a music suppression never goes inside () — (no music) there reads as a music cue (house inference, the default) →
• First/last frames are omni_reference work: the platform start_image/end_image roles or an in-prompt declaration (@Image 1 is the first frame) — which holds better is unmeasured; keyframes 3+ are always in the prompt; never merge two anchors into one sentence →
• Editing needs a sole editing master + edit scope + Timeline Inheritance; extension needs the boundary frame aligned before any new content: MODE-PLAYBOOKS.md
• Storyboard grids, coarse-vs-fine blockouts, one-click video, seamless transitions: MODE-PLAYBOOKS.md
• 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
• Emotion needs 2–4 observable cues, not adjectives; niche camera terms get translated into a visible result →
• 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/Race] label →
• Hard limits that must not be over-promised (frame accuracy, locked parameters, pixel-identical transitions) →
• Dreamina-product features that are not on the Higgsfield surface — Ultra Long Video 180s, mark-based editing, Clay Renderer →

Provenance

Two primary sources, labelled throughout, plus the secondary sources below. What each label means — and the evidence it requires — is the repo-wide legend in ../shared/provenance.md.

Where the two disagree about what is settable, the platform snapshot wins — it is what the API actually accepts.

The Mode Router

Pick the mode first. The same sentence means different things in different modes, and two of the four modes are not generation at all.

Two rules that fall out of this:

1. First/last frames, keyframes, storyboard grids, and blockouts are all omni_reference. 2.5 has no separate first/last-frame mode on this platform. The first and last frame can go in two ways: the platform start_image / end_image roles (allowed only in omni_reference — the catalog rejects them in t2v, video_edit and video_extension), or ordinary references whose role sentence says they are the first and last frame. Which of the two holds the boundary better is unmeasured; keyframes 3+ exist only in the prompt form. [OFFICIAL — Dreamina: "no need to switch to a separate first/last-frame mode"] · [OFFICIAL — platform, snapshot 2026-09-26]
2. Editing is not regeneration. If the user wants the shot rebuilt, that is omni_reference with the old clip as a motion reference — not video_edit. video_edit preserves the master's timeline and changes one scoped thing inside it.

Video-to-video is not automatically video_edit. The field VFX workflow — swap the person in this plate, keep every other pixel — runs in omni_reference with the source attached as a video reference, because that is the lane where duration is settable and must be matched to the source (and where the source therefore has to be ≥ 4 s, the duration floor). video_edit is the lane for a scoped change inside a master whose timeline must survive untouched. Full routing table + the performance-inheritance clause: VFX-PIPELINE.md § Stage 4. [FIELD — AI-vs-VFX, 2026-08-08]

The Higgsfield Parameter Surface

[OFFICIAL — platform, snapshot 2026-09-26] · verify against ../../specs/model-specs.json before quoting (HARD RULE 3).

The catalog also carries reference-count rules that the table cannot show [OFFICIAL — platform, CLI rules 2026-09-26]:

• t2v takes zero references — no image/video/audio references and no start/end frame.
• omni_reference needs at least one reference (a start or end frame counts).
• start_image / end_image are accepted only in omni_reference.
• image references + start frame + end frame ≤ 30; all references + start + end ≤ 50.

Three consequences worth stating to the user before they spend credits:

• video_edit bills by the source video's duration, and neither duration nor aspect_ratio is settable — a 20-second master costs a 20-second render no matter how small the edit.
• video_extension inherits the source's aspect ratio; only the extension's duration is yours to set.
• No genre parameter. 2.0's genre hint does not exist here — genre lives in the prompt's visual-style clause instead.

Preflight the same way as 2.0:

python3 scripts/seedance_lint.py --preflight --model seedance_2_5 "<prompt>"

The linter reads the enums out of ../../specs/model-specs.json, so an out-of-range duration, a 4K request, or a video_extension missing its extension_mode is caught before the render.

The Core Prompt Formula

[OFFICIAL — Dreamina] Combine only the parts the shot needs; omit the rest rather than padding empty slots.

<Subject> performs <primary action or event> in <scene and environment>.
The visuals feature <visual style>.
Use <shot size, camera angle, camera movement, or cuts>.
Audio includes <dialogue, ambience, sound effects, or music>.

• Subject + action is load-bearing — make it concrete. "The man runs" → "the man accelerates into a sprint while his jacket reacts to the airflow."
• Scene and environment — location, time, weather, spatial relationships, background state.
• Visual style — lighting, color, materials, texture, mood. Only descriptors that add information; stacked buzzwords ("cinematic, 8K, masterpiece") sample nothing in particular. The named-substitute discipline in ../higgsfield-seedance/SKILL.md § Prompt-Craft Laws applies unchanged.
• Camera — shot size, angle, movement, focus subject, transitions. Motion matches the action; it is not decoration.
• Audio — dialogue, voice characteristics, ambience, SFX, music, synchronized to the visuals.

Generation parameters are not prompt text. Resolution, duration, and aspect ratio are set on the generation page or via the API — writing them into the prose does nothing except in the modes that auto-lock them, where they are not settable at all.

Reference Roles — Say What to Use and What Not to Use

The moment there is more than one material, or a material sitting next to a text description, the prompt must state what each material contributes. [OFFICIAL — Dreamina]

@Image 1 defines <subject>'s <appearance, clothing, structure, or material>.
@Video 1 defines <motion, camera movement, or pacing>.
@Audio 1 defines <character or sound type>'s <voice, dialogue, ambience, or music>.

Every material that could leak something unwanted gets an explicit exclusion in the same sentence:

@Image 2 defines the workbench and window light. Do not use the people in the image.

Rules:

• Mappings live in the prompt. Text labels drawn inside an image are not a mapping, and the model will not infer which person or prop a material represents.
• Video-only references are motion/pacing references by default, not identity — say so explicitly when you mean otherwise.
• Several views of one subject must say they are one subject: "All four images define one folding desk lamp. The output must contain only one lamp throughout." Without it the model duplicates the subject.
• When a reference video already carries the motion accurately, state only which attributes to inherit. Restating every action fights the reference. A blockout or motion video carries motion and spatial structure — not identity — so the prompt still has to define subjects, scene, action, and visual style.
• Never place a reference handle in a shot where that subject is absent — the same rule as 2.0's tag discipline; the model forces it into frame.
• Character sheets leak their staging. A sheet's neutral backdrop and multi-view panel layout are the most common character-material leak — the flat gray studio renders as the actual set. Pair every character-sheet role with its own exclusion: "Do not take the gray backdrop, the panel borders, or the multi-view layout."
• Beat lines name characters, never handles. In action/beat prose, a character appears as name + one visible marker at their first appearance in the beat — "Mira — silver streak, rust-red jacket — crosses the stall line" — not as @Image 2. The model binds by what it can see in the material, and a handle used as a sentence subject is the classic way one character comes back as two people. [EMPIRICAL — sd25-pe mapping priority, re-derived 2026-08-09: material content outranks upload order] Scope: this is the 2.5 rule. On 2.0 the house convention is the opposite — the acting paragraph leads with the character's tag so the model binds the performance to the right person (../higgsfield-acting/SKILL.md § Scene adaptation, ../higgsfield-seedance/SKILL.md § Tag naming). Handles stay in the role map on 2.5 — the [Characters] lines, a staging legend (@A = the BLUE figure) — and out of the beat prose (../shared/house-rulings.md P2-13). Handle spelling follows the surface: Dreamina's guide writes upload-order handles with a space (@Image 1); the Higgsfield field build writes the upload-order form without one (@video1) and named asset tags (@size-ref, @dragon-v3). Pick one form per project and never mix them in one prompt.

Fidelity — say how much of each material must survive

[EMPIRICAL — MiniMax H3 skill corpus, re-derived 2026-08-09; cross-model structure, unmeasured on Seedance] A role says what job a material does; it still doesn't say how much of the material must reach the pixels. Declare one fidelity grade per material, in the same sentence as its role:

• full-preserve — the subject appears as-is: face, build, wardrobe, all of it.
• partial-preserve — the named parts survive, the rest is free: "the jacket and the scar; hairstyle may change."
• attribute-transfer — named traits lift onto a different, named target: "apply this fabric's weave and sheen to Mira's coat." The target must be named — this is the one case a bare role line cannot express, and the one that goes wrong silently.
• loose-guide — mood, palette, or energy only; nothing is copied literally.

@Image 4 defines the brocade fabric — attribute-transfer onto Mira's coat only.
Do not carry the garment's cut, the mannequin, or the studio backdrop.

Material budget

[OFFICIAL — Dreamina] Hard limits vs the ranges that actually stay stable:

50 reference materials total. On Higgsfield the platform enforces two of these caps and counts a start_image / end_image against both: image refs + start + end ≤ 30, all materials + start + end ≤ 50 [OFFICIAL — platform, CLI rules 2026-09-26]. The 10-video, 10-audio and ≤30 s-combined figures are Dreamina's; no platform rule states them — treat them as the vendor's model limits, unverified on Higgsfield. Above the stable ranges (9–12 subjects in images, 6–10 in audio/video, 6–8 edit reference images) generation still works but stability drops and the shot may need several attempts — budget for it, or split the scene.

More than five subjects needing multiple views → one view per image. Independent view images beat a single collage of views; the collage is the less stable form.

Spend one view on a strong expression, not four resting faces. [OFFICIAL — Higgsfield Seedance 2.5 deck, PART 2] A set of neutral views teaches the model the face at rest and nothing else, so the first line of dialogue invents a mouth. Generate the views on a neutral light-grey ground (shade and mechanism: ../../templates/ad-asset-prep.md § Design for win rate) and make one of them a strong expression — anger, or a wide smile — so the model learns the character's facial dynamics and teeth structure, not only the resting face. The canonical four: front view · back view · facial details at neutral · facial dynamics and teeth under strong emotion. Close the set with the identity line (§ Reference Roles) so all four are read as one person.

Multi-Reference — the Five-Step Workflow

[OFFICIAL — Dreamina] The goal is not to cram every reference into one sentence. It is to define the relationships among characters, props, scenes, actions, and audio so the model picks the right material for the right moment.

Step 1 — name and map each subject individually. One line per subject:

<Character A> corresponds to @Image 1. Use only the appearance, hairstyle, and clothing.
<Character B> corresponds to @Image 2. Use only the appearance, hairstyle, and clothing.
<Prop A> corresponds to @Image 3. Use only the structure, material, and color.
<Scene A> references @Image 4. Use only the spatial layout, architecture, and lighting.
Do not use the people in the image.

The canonical failure: @Images 1 through 4 define four characters respectively. That sentence does not say which image is which character, and the model will guess.

Step 2 — group by type once several subjects are in play: [Characters] → [Props] → [Scenes] → [Motion and Audio]. Add the non-interchange lock to the character group: "Do not interchange these characters' appearances, clothing, actions, positions, or dialogue."

Step 3 — profile any recurring subject. A character crossing several scenes, or carrying several materials, gets one consolidated block:

[Subject Profile: Conservator]
Appearance and clothing: @Image 1.
Fixed prop: <Sample Case> from @Image 5.
Locations: <Conservation Lab> and <Gallery>.
Motion references: the case-opening motion from @Video 1.
Do not use: other characters' clothing. Do not give this character other equipment.

Step 4 — select references by scene. Per scene, list only the subset actually used, then the event and its end state:

Scene 1 | Inspection in the Conservation Lab
Use: <Conservator>, <Sample Case>, <Conservation Lab>, and the case-opening motion from @Video 1.
Event: <Conservator> opens <Sample Case> at the workbench and inspects the sample inside.
End state: <Conservator> remains on the inner side of the workbench. <Sample Case> stays
beside the conservator's right hand.

Step 5 — check ownership. Props belong to exactly one character ("belongs only to <Conservator>"); character count, clothing, and spatial direction stay constant across scenes.

Long Video — Stages and End States

[OFFICIAL — Dreamina] 2.5 generates up to 30 seconds natively. Anything with several events gets staged — one flat paragraph is where dropped beats come from.

Each stage carries exactly one primary state change and closes on an explicit, directly visible end state. Each new stage restates what carries over.

[Generation Goal]
Generate a <video type>. The central subject is <subject>, and the primary event is <story summary>.

[Stage 1]
Initial state: <initial state of characters, props, and scene>.
Primary event: <one primary action or event>.
End state: <character positions, prop ownership, or visible scene state>.

[Stage 2]
Continue from the previous stage: <state that must remain unchanged>.
Primary event: <one primary action or event>.
End state: <observable state>.

[Stage 3]
Primary event: <closing event>.
End state: <final visible state>.

[Maintain Consistency]
Keep <character identity, number of characters, clothing, prop ownership, spatial direction,
and audio relationships> consistent.

Split by JOB, not only by length

Staging solves too many events in one paragraph. It does not solve two incompatible jobs in one generation, and that is a separate cut worth making [DEMO — Higgsfield "AI Love Stories" tutorial, 2026-08].

The reported case: an arena scene containing a fight (a giant throw, a sword snatch, a takedown) and, in the same beat, the acting around it (a hidden worry, a flirtatious exit, hope draining from a face). Run as one prompt it held — badly, in both directions at once: "the fight gets softer, the faces get flatter." The model spends its attention budget once. Asked to nail physics and micro-performance in one generation, it half-does each.

Split into two prompts, one job each, stitched in the edit:

How to tell which cut you need. Length-driven splits are decided by counting events. This one is decided by asking: what is this shot's dominant job? If a prompt has two honest answers, it has two shots in it. That is the same question the Feasibility Veto asks about a single frame, applied to a whole generation.

Note the direction of the second case: it fit, and was split anyway. "It generated" is not the bar — a scene that renders every beat but rushes the one that matters has failed at the thing you were making it for.

Practical consequence, stated plainly by the same production: a perfect 30s render does not exist. Generate raw footage per job, then assemble. Several of the finished scenes in that film take the background from one take and the foreground action from another.

Timestamps and pacing

Stages are the default. Reach for one-second precision only when a specific handoff, entrance/exit, transition, or beat must land on a moment.

• Ranges are consecutive and non-overlapping, and they are a time budget, not a frame-accurate edit point — actions may land slightly either side of a boundary.
• Too little content in a range hands the model freedom you did not intend; too much causes excessive cutting or dropped events.
• Never demand a frequency inside one second ("complete three actions in one second"). Timestamps allocate time; they do not enforce cadence.
• Every timestamped segment still ends on an explicit end state, exactly like a stage.

Relationship to 2.0's beat arithmetic. ../higgsfield-seedance/SKILL.md § Output Format requires timed beats to sum to the declared duration. That still holds — but sum to 4–30s here, and remember that on 2.5 the sum is a budget the model approximates, not a cut list it honours to the frame.