higgsfield-seedance

Higgsfield Seedance Director

QUICK FACTS

Generated-checked block (scripts/build_index.py verifies anchors). Read the linked sections for full context — these lines are routing aids, not the rules themselves.

• The filter is an LLM reading full-scene intent, not a keyword blacklist — describe a SCENE, not a subject; fix the voice first →
• Instant fail (<10s) = filter rejection; delayed fail (>30s) = infra/complexity — never regenerate an instant fail unchanged →
• Six slots, in order: Camera + Subject + Action + Setting + Style + Lighting; missing 3+ slots is where flags come from →
• Empirical prompt-craft laws: 50–80-word attention sweet spot (front-load the load-bearing element), name a director/lens not "cinematic", "fast" degrades motion, no negative prompts in the body →
• Five prompt modes: Reference-Based / Continuation / Expand Shot / Edit Shot / Transformation — pick the mode before writing →
• [OFFICIAL] block scaffold for production prompts: SCENE CONTEXT → … → POSITIVE LOCKS, distributed style (no prefix), FOV in degrees only, CAMERA block 3rd, cut ladder oner / CUT n / timed / freestyle →
• [FIELD] 13-project corpus calibration: word length scales with register (218w → 2,059w medians — the 50–80w sweet spot is single-shot-only), video briefs hand-authored (enhance_prompt off), Style Prefix = per-project constant compiled into home blocks →
• [FIELD] Three "helpful-instinct" drift sources, each with a standing lock: environment invention (#1, above character drift), character-height equalization, scale drift on wides →
• Build-safe construction for crowds/destruction/creatures: evacuated cities, contained fights ("stays at the sea surface"), the safe benchmark scene →
• Extend an existing clip: attach it as a video reference + open with "The scene continues." — match source resolution AND duration; chain cap ~2 (hard 3), then re-anchor from ORIGINAL references →
• Tutorial-demonstrated patterns (reference-role vocabulary incl. VARIETY reference, SCREEN REALISM + duration-match composites, 60:30:10 grade, red-arrow prop annotation): PRODUCTION-PATTERNS.md in this directory
• Hard engine rules (age-blind, exit-frame = cut, off-screen = nonexistent, no reflections, ≤3 tracked characters, double-contrast cuts) + high-risk shot table: ENGINE-RULES.md in this directory
• Reference roles: Character / Last-Frame / Environment / Prop — role determines what the prompt may re-describe →
• Working modes: Exploration / Continuation / Bridging / Repair (distinct from prompt modes) →
• Layer 1 briefing vs Layer 2 production prompt — never paste Layer 1 into the prompt box →
• Native 4K is available in mode=std only; mode=fast (Seedance 2.0 Fast) caps at 480p/720p — in Cinema Studio the model is still capped at 1080p →
• Always preflight: python3 scripts/seedance_lint.py --preflight --model seedance_2_0 "<prompt>" — enums come from ../../specs/model-specs.json (fast+1080p/4K and Kling 21:9 are auto-caught) →
• 480p drafts validate the prompt, NOT the take — no seed param; pin Hero Frame + start/end frames to carry a look →
• ZH prompts: hard 1,800-char cap; ZH antislop list enforced by the linter →
• Flagged prompt → rewrite playbook per linter rule, then voice pass →
• Repeated flags → full loop-breaker procedure + LOG THE OUTCOME (--confirmed / add-quality) →

Use this skill whenever the user wants a Seedance 2.0 / Seedance Pro prompt, OR whenever a Seedance generation has been blocked, flagged, or silently failed. This skill's job is to stop credit waste on filter rejections.

Engine rules (read with this file): the hard rendering constraints of the Seedance 2.0 engine — age-blind characters, exit-frame = implicit cut, off-screen = nonexistent, no reflection shots, ≤3 tracked characters, the double-contrast cut rule — live in ENGINE-RULES.md in this directory, together with the high-risk shot table (reflections, same-character doubles, crowds, text rendering) and its mitigations. This SKILL.md is the EN-director profile of that rule core; the ZH-house and bilingual-JSON profiles (../../docs/Seedance 2 Skill.md) obey the same core. Flag high-risk shot types at authoring time — never silently break a rule the project's hero image happens to conflict with.

Production patterns (sibling reference): patterns demonstrated working in Higgsfield's own Seedance-4K film tutorial — reference-role vocabulary, coordinate blocking, video-reference screen composites, prompted imperfection, 60:30:10 grade — live in PRODUCTION-PATTERNS.md in this directory, labeled [DEMO].

The Filter Model — Read This First

Seedance 2.0's content filter is not a keyword blacklist. It is a language model that reads the full prompt as a single scene and judges intent and context. Most users burn hours swapping individual words — that loop does not work.

The filter compares two things:

• A prompt that reads like a filmmaker describing a shot → tends to pass.
• A prompt that reads like a note to a friend → tends to fail.

A word that looks sensitive in isolation can sit inside a well-constructed cinematic prompt without issue — the filter reads the full picture. A prompt with no picture to read (no setting, no visual purpose, no narrative logic) gives the filter nothing to work with, and it errs on the side of caution.

Practical rule: the prompt must describe a scene, not a subject. Fix the voice first, then fix the words.

Instant Fail vs. Delayed Fail — the Diagnostic

This single heuristic saves time on every failure:

If the user is seeing instant fails in a loop, it is a filter issue — never a GPU issue. Stop them from regenerating before the rewrite.

The Seedance Prompt Formula

Every Seedance prompt should hit these six slots, in this order:

[Camera movement] + [Subject] + [Action] + [Setting] + [Style] + [Lighting]

All six are technically optional — but a prompt that includes all six almost never gets flagged, because the filter has full context to interpret every word. A prompt missing 3+ slots is where flags come from.

Minimum viable Seedance prompt

Slow dolly-in on a figure in a dark overcoat standing alone at the end of a rain-slick alley. Cold teal shadows, single practical streetlamp, shallow depth of field.

Camera ✓ Subject ✓ Action ✓ Setting ✓ Style ✓ Lighting ✓ — all six slots, ~30 words, passes the filter because the scene is fully legible.

Prompt-Craft Laws

A set of Seedance-2.0-specific prompt rules. These are empirical — practitioner A/B findings that are plausible given the architecture but are not in the official model spec. Treat them as strong heuristics and let the repo's iteration discipline (../higgsfield-prompt/SKILL.md § The Iteration Rule) confirm them on your own material, rather than as guaranteed model behavior.

Length and order — the attention model

Seedance reads the prompt left-to-right with diminishing attention weight. The first sentence carries the most influence; by the third sentence you are in "detail territory," where the model stops treating elements as primary instructions and starts sampling them diffusely.

• Sweet spot: 50–80 words. A 70-word prompt reliably outperforms a structurally identical 200-word version of the same scene — more words past ~3 sentences buys diffusion, not control.
• Structure in three sentences: ① subject + action, ② camera + style, ③ constraints / positive locks.
• Lead with the single most load-bearing element. When a shot lives on its subject, the subject opens the prompt; when it lives on a camera move, the move opens it.

Relationship to the two length numbers. This 50–80-word figure is the coherence optimum. The >180-word figure in § Pre-flight Linter is a different axis — the filter/encoder risk ceiling (>220 often hard-fails the text encoder). 50–80 is where to sit; ~180 is where it starts to break. They don't conflict.

Relationship to the six-slot formula. The six slots guarantee the filter sees a complete scene (presence). The attention model governs weight (order

• length). Keep all six slots present, but the slot list's camera-first ordering is a completeness checklist, not a mandate to open with the camera word when the shot's identity is the subject.

Name the thing — kill empty adjectives

cinematic, epic, beautiful, high quality, amazing are high-frequency labels attached to an enormous range of training footage — dark thrillers, bright rom-coms, nature docs all read as "cinematic" — so the model samples a broad, diffuse distribution and they move the output toward nothing in particular. Don't just delete the slop word (Voice Rewrite §6) — substitute a named, narrowly-trained referent:

| Empty adjective | Named

Content truncated.