Higgsfield Seedance VFX — Footage Transformation

This skill is for editing a clip the user already has: keep a real subject and the real camera move, change only what they ask for. It is the video-to-video sibling of ../higgsfield-seedance/SKILL.md (the general Seedance director). It reuses that skill's grammar verbatim — the six-slot formula, the Prompt-Craft Laws, the preflight linter — and adds the transformation layer below. Do not contradict the parent skill; when a shot needs filter-safety, mode selection, or engine rules, defer to it.

> This file is the Seedance 2.0 v2v lane. It stays the right choice when the job needs > 4K, mode=std, a platform start/end frame, or a genre hint. If the plate job is > running on Seedance 2.5 — 720p ceiling, up to 30 s, many references — the lane is > omni_reference with the clip attached as a video reference, and its production doctrine > (source ≥ 4 s, duration = source, the four-batch rule, the performance-inheritance clause) > lives in ../higgsfield-seedance-2-5/VFX-PIPELINE.md.

> This skill is a video-to-video layer on top of ../higgsfield-seedance/SKILL.md. > Every VFX transform is still a Seedance prompt. This skill only changes the starting > point: a real source clip whose subject, performance, and camera move must survive the > edit. higgsfield-seedance § Seedance 2.0 Prompt Modes / Transformation covers the > in-clip morph (a character visibly becoming something else in one take); this skill > covers preservation-VFX — lock the real plate, layer the effect in.

QUICK FACTS

Routing aids — read the linked sections for the actual rules.

• The one job: preserve everything that makes the source recognizable, change only the named element; repeat the fragile guardrail ("face and identity unchanged") at the end →
• Run it in Seedance 2.0, mode std, 4K — faces, lip-sync and fine detail hold at 4K where they warp at 1080p (fast can't do 4K; Cinema Studio caps at 1080p) →
• @source declares the clip as the base (not a style ref); add @creature/@element only when a real texture keeps getting faked →
• Three levels of difficulty: L1 swap the world · L2 change an element in-frame · L3 full handheld cinematic →
• Two modes: add an element to the plate, or replace the environment around a preserved subject →
• Color matching alone looks pasted-in — match key direction, bounce, optics/haze, edges/grounding →
• Photoreal creatures need biological accuracy (wrinkled, cracked, asymmetric, matte — never smooth/glossy/inflated) + a real contact shadow; a reference image beats a description →
• Timed zoom synced to a line: anchor it twice — semantic (On the line "…") + numeric (At about Ts); see references/dialogue-timing.md →
• Prepended-intro budget: total − intro = surviving window for the source performance; recompute on every change →
• Generate the transformed start frame first to lock the look before spending video credits; see references/first-frame.md →
• Output is plain-text English, no markdown inside the prompt, easy to copy →

The core idea: preserve, then change one thing

A transformation prompt has two jobs that pull against each other: lock everything that makes the source recognizable (the person's identity, face, wardrobe, performance, framing, lens, and camera motion), and change only the named element. If you under-specify the lock, Seedance re-rolls the face or the camera and the edit stops matching the original. So every prompt states both halves explicitly, and repeats the most fragile guardrail — usually "face and identity unchanged" — at the end of the action.

This is the whole blueprint: real footage goes in, the exact same shot comes back — same subject, same movement — with only the one requested change applied.

Resolution: run it in 4K

Run everything in Seedance 2.0, mode=std, 4K. The 4K matters: faces, lip-sync and fine detail hold at 4K where they warp and fall apart at 1080p — and a footage transform lives or dies on the preserved face reading as the same face. The 4k resolution enum is model-verified (../../specs/model-specs.json → seedance_2_0); the "detail holds at 4K" observation is a practitioner claim, strong but not a spec guarantee.

Two model constraints to respect (from ../higgsfield-seedance/SKILL.md § Pre-flight Linter): native 4K is available in mode=std only — mode=fast (Seedance 2.0 Fast) caps at 480p/720p, and inside Cinema Studio the model is still capped at 1080p. So a 4K footage transform must run in std mode, on the standard Seedance 2.0 surface. If the user is on fast mode or in Cinema Studio and asks for 4K, flag the cap before they generate.

Prompt anatomy (transform variant)

1. The @source declaration

The source clip is the base, not a style reference. One line:

@source: Original <clip name> — <who/what is in it: subject, wardrobe, setting, action>. Preserve
<identity, face, wardrobe, performance, framing, camera and motion> exactly; <what to change —
e.g. enhance only the environment / add the creature on the tower / transform only the right arm>.

If a transformation needs a real texture the model keeps faking (an animal's fur, a specific face), add a second input as a texture reference and declare it:

@creature: Reference photo of a real <animal> — <fur / face / anatomy notes>. Appearance and
fur/skin texture reference only; ignore the photo's background and lighting, do not use it for the
environment.

The user supplies their own descriptions of what is in their files — use the tags correctly, don't invent what the clip contains. But before writing @source for a clip you can open, inspect it: read its duration / fps / aspect and extract a few frames. Build @source and the specs runtime from what the footage actually shows — subject, wardrobe, framing, camera move, time of day, key direction — not from the user's one-line summary. Set the specs duration to the probed runtime by default. If no source clip is described, ask what footage they're starting from before writing.

> Seedance reads an uploaded video as a set of frames, not as a watched clip — so the frame > content (subject, walk, light) is the context it has. Same principle as the screenshot → > Claude workflow in ../higgsfield-seedance/SKILL.md § Load-Bearing Rule.

2. Specs line

One compact line. Always include the source-matching constraints:

Photoreal. <aspect, default 16:9>. <duration — match the source clip>s. 4K. <look / grade>.
NON-IP — generic <creature/design>, not based on any brand or character. <SFX only | SFX and
source dialogue only>.

• Match the source runtime by default. If the clip is 6s, the prompt is 6s. Extend only

when a payoff needs room (a slow creature turning to camera), and say why.

• NON-IP guardrail belongs in the specs line whenever a creature, armor, vehicle, or

character design is added — generic, never a branded character. Keeps outputs clean and tends to generate more reliably than a trademarked design. (Same filter logic as ../higgsfield-seedance/SKILL.md § The Rewrite Playbook / Brand · IP.)

• Audio: SFX only for added effects; SFX and source dialogue only when the source talk

track must survive (e.g. a zoom synced to a spoken line — see references/dialogue-timing.md).

3. Scene action — one continuous shot

The source is a single take, so describe continuous camera movement, not cuts. Lead with the shot/lens and "same framing as the source," then the preserved performance, then the transformation, then any timed camera move. Close with the lock-down clause.

4. SFX line

End with a specific, ordered SFX note, exactly as the parent grammar requires. For added effects be behavioral: not "fire" but "a soft whoomph as it catches, then a low steady flame roar and crackle, occasional ember pop." Sync every effect (footsteps, impacts, wind, creature calls, servo whirs) to the visible action.

Three levels

The workflow scales in difficulty. The steps are identical at every level — inspect the clip, name the one change, lock the rest — but the harder the camera moves, the harder the effect has to track:

• Level 1 — swap the world around you. Keep the subject exactly as shot; replace the

environment. Easiest on slow/steady moves; a driving shot raises the bar because the new world has to stream past at the right speed and relight the subject as it goes (§ Two transformation modes / B).

• Level 2 — change an element in the frame. Add or morph one specific thing: set a head of

hair on fire, morph a hand into something unexpected, add a creature climbing a building. The plate stays; the effect is layered in and lit into it (§ Two transformation modes / A).

• Level 3 — full handheld cinematic. The camera is in-hand and moving the whole time — angle,

parallax and shake all changing — and the effect (creature, environment, weather) must track to all of it without falling apart. This is the hardest case; preserve the handheld move frame-for-frame and lean on § Photoreal creature / element integration and § Lighting integration hardest here.

Locked-off shots are easy — the frame barely changes, so the model just paints the effect in. The difficulty is entirely in how much the camera moves.

Two transformation modes

A. Add an element to the footage

Set a head on fire, transform a hand, make a limb invisible, perch a creature on a landmark. Keep the whole plate; layer the effect in.

• Describe the effect's physics and behavior over time, not just its presence: where it

starts, how it spreads, how it moves, what light it throws. Use directional "creep" for transformations ("starts at the tattoo, fine seams split one at a time and peel back, a servo seats, a cable plugs in, the next plate locks…") and ignition-then-build for fire.

• Make the effect interact with the plate: firelight flickering on a face and spilling onto a

car's paint; a glassy invisible arm refracting the background; a giant creature casting a real soft-edged contact shadow on the structure it grips. (The model spills an added fire's orange glow onto skin, shirt and car on its own — but naming the interaction makes it reliable.)

• Scale must be explicit for giant creatures, or the model renders them life-size. Say

"enormous, its massive body dwarfing the structure, clearly colossal relative to the mast."

• The subject usually stays oblivious / unfazed, mid-delivery — that contrast is the joke.

State it.

B. Replace the environment around a preserved subject

Keep the person, their vehicle, the seatbelt, the camera rig and its move; swap the whole world.

• The new world must stream past with parallax consistent with the original motion. If the car

was driving, the replacement must give it a surface to drive on and things that rush past at speed. On a fast move the environment can't just sit behind the subject — it has to move with the subject at the right speed and change the lighting as it goes.

• The bigger the move, the longer the lock list: it's not just the face anymore — it's the car,

the seatbelt, the rig framing, the whole driving motion, all held while only the environment changes.

• Warm, directional daylight worlds are safer for face/identity consistency than night or neon

— those force a full relight of the subject and raise drift risk. Flag this tradeoff and bake the relight instruction in when the user wants night/neon anyway. (Seedance will pull light from the generated world and bounce it onto the subject — neon sliding across a car, lava glowing under a chin — without being asked; naming it makes it dependable.)

Lighting integration (the part that makes or breaks it)

First decide the fork with the user — it changes everything:

• Preserve the subject's lighting, grade only the new elements. Lock the subject's original

light; light and grade the added creature/environment to match the existing key on the subject so they integrate. Lowest identity risk.

• Relight the whole frame under one look. Subject included. Use this for a unified cinematic /

commercial grade. Higher risk to the face, so keep identity/expression/wardrobe explicitly locked while only lighting and grade change.

Color matching alone is not enough to make a preserved subject sit in a new world — that's the most common "looks pasted in" failure. When integrating a subject (or a creature) into a plate, go beyond color with this recipe:

• Light: same key direction (name it — screen-left or screen-right), same softness, same shadow

density and direction across the subject. ("Keep the sun as the key from screen-left exactly as before so the face and the light on it barely change.")

• Environmental bounce: let the world spill onto the subject — cool skylight from above, a warm

bounce from sunlit ground/foliage, subtle ambient occlusion where forms meet.

• Optics & atmosphere: match lens character and micro-contrast; add a touch of the scene's

atmospheric haze and aerial perspective over the subject so they aren't unnaturally crisp against a hazy background; match depth of field, focus falloff and film grain to the rest of the frame.

• Edges & grounding: remove hard cut-out edges, halos and mismatched rims; ground the subject

with believable depth so they occupy the same space.

State the time of day and key direction concretely ("soft, diffused midday daylight with the key coming from screen-right"). "Softer" means a larger, more diffuse source: gentle soft-edged shadows, low contrast, smooth highlight rolloff, light haze. Full cinematic-lighting vocabulary is in ../../vocab.md § Lighting.

Photoreal creature / element integration

When a creature or hard-surface element is added and must read as real:

• Demand wildlife-documentary / practical realism explicitly: "fully photoreal, real fur with depth

and individual strands (or true scale detail / brushed metal), true anatomy, never CG, plastic or cartoonish."

• Biological accuracy is what separates a living thing from a cheap render. Specify real

anatomical detail — "pebbled scaly skin, long claws, heavy tail, true reptile anatomy" — and demand imperfection: deeply wrinkled, cracked, sagging, asymmetric, mud-caked and matte. Avoid smooth, glossy or inflated surfaces; real animals are uneven. Match behavior to the species (a sloth shifts slow heavy weight; a chimp is alert and twitchy; a snake's coils tighten and a forked tongue tastes the air — and snakes don't blink, so use an unblinking stare, not a blink, for a reptile payoff).

• Sell the scale. Use on-screen reference objects (trees, buildings, crew, the subject) to

establish true size, and use a telephoto-lens illusion with shallow depth of field to veil a giant creature partly behind foreground elements — plus a little motion blur and handheld softness so it integrates rather than floats.

• Tie it into the plate: same sun direction and color temperature as the subject, a real

soft-edged contact shadow on what it touches, the same hazy atmosphere and depth as the far background.

• If it still reads as CG after a take, the reliable fix is a second input — a reference photo

of the real animal/material, declared as a texture-only reference (see @creature above). When you already know exactly what you want, it's better to show the model than to describe it; generate the reference in ../higgsfield-gpt-image-2/SKILL.md (GPT Image 2) or Nano Banana Pro, then point the prompt at it.

Timed camera moves synced to dialogue

A crash zoom or smooth push-in landing on a beat is a recurring payoff. Anchor it two ways at once so it lands even if Seedance's internal timing drifts: a semantic cue and a numeric cue. The full measurement procedure — how to read T off the source audio and convert a timecode — is in references/dialogue-timing.md.

• Semantic: On the line "<exact words>," the camera <snaps into a hard crash zoom | begins a

smooth, steady push-in> …. Requires SFX and source dialogue only in the specs so the talk track survives.

• Numeric: At about <T> seconds … the camera …. Get T from the source audio (or a visible

action beat — "at about 2.2s, on his finger snap").

• Crash zoom = fast hard punch-in; smooth push-in = slow steady glide, no snap. Match the

user's word.

• If a landmark or subject must stay visible through the move, say so explicitly ("the tower

stays in frame throughout, never cropped").

• Leave enough tail after the trigger for the payoff to play (a creature slowly turning to camera

needs ~2–3s). If the clip is short, fire the zoom on the first word of the line rather than after it.

Reveal pull-back (the outward move)

The mirror of the push-in: open tight on the added element in isolation — a long-telephoto, compressed framing of the creature/effect with the subject out of frame — then move outward to land on the real plate. Two flavors, match the user's word:

• Hard / snap zoom-out = a fast punch outward, abrupt.
• Smooth pull-back = a slow steady decompression, no snap.

Anchor the landing the same two ways as a timed zoom, and demand a 100% match of the source composition at the landing: name the matched attributes — same angle, headroom, horizon, lens character — or the model lands on a near-miss framing that no longer cuts against the original. After the landing, hand off to the preserved take and keep the source's own camera motion running.

Preserving lip-sync to a known line

When the payoff is the subject's mouth matching a specific line, quote it verbatim and anchor it twice: once inside the action ("…lips matching the source exactly, saying clearly: '<line>'…") and once in the SFX/dialogue line. Require SFX and source dialogue only in the specs so the talk track survives, and add "lips matching the source exactly" to the lock-down clause. Then check the line against the surviving dialogue window (see § Duration discipline / Prepended-intro budget) — a line that runs ~6s cannot sit in a 5s tail. If it doesn't fit, resolve the runtime before delivering; don't ship a prompt that can't lip-sync. Dialogue + [AUDIO: Xs] mechanics live in ../higgsfield-audio/SKILL.md.

Duration discipline

Default to the source clip's exact runtime. When the user changes the runtime, recompute any numeric zoom timing and tell them the new mark. When a long hold lands on a static creature, add small "living" micro-movements (a slow blink, jaw shift, steady breath) so it doesn't look frozen.

Prepended-intro budget: intro + remaining = total

When you prepend a beat (a reveal, a telephoto hold, an establishing creature shot) to footage you must preserve, the preserved take does not get longer — it gets pushed back. State the arithmetic every time and flag what falls off:

total runtime − intro length = surviving window for the source performance

If the source take is longer than that surviving window, some of it cannot play. Say so explicitly and offer the three resolutions, in order of fidelity:

1. Extend the total so the full source fits (intro + full source). Highest fidelity, longest

clip.

1. Start the source earlier — sacrifice the clip's own quiet lead-in so the dialogue still

lands in the window. Keeps total fixed, keeps the words, loses pre-roll.

1. Accept truncation — the first N seconds of the source won't appear. Only safe if the dropped

head has no dialogue.

Never promise "100% lip-sync" and a prepended intro on a fixed total without doing this subtraction first. Recompute and re-flag it on every change to either number.

First / start-frame workflow

Before spending video credits, it's often worth generating the transformed opening still as an image, locking the look, then animating from it. See references/first-frame.md for the full procedure (model, settings, inputs, upload mechanics, and how to hand the still back to Seedance as a start_image). This pairs with ../higgsfield-seedance/SKILL.md § Drafts Validate the Prompt — pin the frame, not the roll.

Iterating

The user iterates fast and in small steps ("softer light," "from the right," "bigger snowier mountains," "make the chimp huge," "a beat before the zoom," "keep the original runtime"). Change only the named thing and keep the rest of the prompt stable — re-rolling the whole prompt loses what already worked. When refining a generated still, edit the chosen result (pass it back as the base) and fix only what's off rather than starting over. This is the footage-transform case of ../higgsfield-prompt/SKILL.md § The Iteration Rule.

Output format

Output in English first, plain text — no bold, no headers, no bullets inside the prompt, not in a code block. Easy to copy as-is. Chinese translation only if asked, after the English, same format.

A short label above each prompt (e.g. Hook_2 · Variant 1 — Through the clouds) is fine and helps when you deliver several variants; the prompt body itself stays plain text.

Skeleton:

@source: ...
@creature: ...            (only if a texture reference is used)

Photoreal. 16:9. <N>s. 4K. <look/grade>. NON-IP — generic <X>. SFX [and source dialogue] only.

<Continuous shot, same framing as source. Preserved performance. The transformation, with physics
and plate interaction. Any timed camera move with semantic + numeric anchor. Lock-down clause: face
and identity unchanged; everything else identical to the source.>

SFX [and source dialogue] only: <specific, ordered sounds>.

Voice (match the user)

Terse and kinetic; physically precise (exact materials, behaviors, scale); director-minded (lenses, angles, moves); non-generic (no "beautiful / stunning / amazing" — texture words instead); emotionally controlled. Don't inflate, don't soften, don't explain what things "represent." Same anti-slop discipline as ../higgsfield-seedance/SKILL.md § Voice Rewrite.

Seedance 2.0 input limits (reference)

Images ≤ 9; videos ≤ 3 items, total ≤ 15s; audio ≤ 3 MP3s, total ≤ 15s; total mixed inputs ≤ 12; generation duration 4–15s. A source clip plus a texture-reference photo fits easily. If a request needs more inputs than allowed, flag it and say what to prioritize. (Verified against ../../specs/model-specs.json → seedance_2_0: media_roles include video, image, audio, start_image, end_image; 4k is a legal resolution in mode=std.)

Structure patterns to internalize (style reference only — do not reproduce)

• Add-element: @source (preserve all, add effect) → specs + 4K + NON-IP + SFX only →

continuous shot, preserved performance, effect igniting/creeping with plate interaction, subject unfazed → lock-down clause → SFX.

• Environment-swap: @source (preserve subject + vehicle + rig + motion, replace world) →

specs + 4K + grade for the new world → continuous shot from the same rig, new world streaming past with parallax, relight to match or relight-all → lock-down → SFX.

• Creature-on-landmark with timed zoom: @source + @creature (texture ref) → specs + 4K +

NON-IP + SFX and source dialogue only → continuous locked shot, giant photoreal creature integrated on the landmark, subject delivering to camera → at ~T / on the line "…", smooth push-in keeping the landmark in frame, creature turns to camera → lock-down → SFX and dialogue.

• Prepended reveal intro (transform + outward move + preserved performance): @source (preserve

subject + performance + lip-sync + framing, add element, prepend a telephoto intro) + @creature (texture ref) → specs + 4K + NON-IP + SFX and source dialogue only → open tight/telephoto on the added element in isolation for the intro beat, hard or smooth zoom-out at ~T landing on a 100% match of the source composition, then the preserved take plays with exact lip-sync to the quoted line while the added element continues behind → budget check (intro + remaining = total) → lock-down → SFX and dialogue.

A ready-to-fill skeleton plus worked variants live in ../../templates/seedance/footage-vfx-transform.md.

Related Skills

• ../higgsfield-seedance/SKILL.md — the parent Seedance director: six-slot formula, Prompt-Craft

Laws, engine rules, content-filter preflight linter, and the in-clip Transformation prompt mode

• ../higgsfield-audio/SKILL.md — audio-as-conditioning, [AUDIO: Xs] dialogue/SFX, lip-sync (the

timed-zoom-to-dialogue and source-dialogue-preservation cases)

• ../higgsfield-camera/SKILL.md — video reference, the Load-Bearing Rule, camera-move vocabulary

(handheld preservation + moves you never filmed)

• ../higgsfield-gpt-image-2/SKILL.md — generate the creature/texture reference image and the

transformed start frame

• ../higgsfield-facs/SKILL.md — muscle-level facial control when a dialogue-preserving transform

also needs an exact expression

• references/dialogue-timing.md · references/first-frame.md — the two procedures this skill leans on