Higgsfield Seedance Vfx

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, or a genre hint. If the plate job is running on Seedance 2.5 — up to 1080p, up to 30 s, many references, platform start/end frames in omni_reference — 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.