Chapter 38. Layered Compositing for Complex Shots#

38.1 Compositing is not a patch applied after failure#

When a shot needs stable characters, legible text, weather, two-person contact and effects all at once, layering is usually more reliable than a single generation. Plan layers, occlusion and tracking during shot design — not as a rescue at the end.

38.2 The basic layers#

A common order: background plate, background characters, foreground characters, hands and props, screen text, weather and particles, light effects, depth of field, unified grade and grain. Not every shot needs all of them.

composite_plan:
  shot: E006_S18
  layers:
    - LOC_BANQUET_BLACKOUT_PLATE
    - CH_LINWEI_STAGE
    - CH_LINYUN_FOREGROUND
    - PHONE_SCREEN_HOSTAGE_LIVE
    - EMERGENCY_LIGHT_GLOW
    - CROWD_SILHOUETTES
  tracked_surfaces: [phone_screen]
  occlusion: Lin Yun's thumb covers the lower-right corner of the screen
  depth_order: crowd < linwei < linyun < phone

38.3 Compositing two people#

Generating separately prevents face bleed, and it requires matching camera height, perspective, focal length, light sources, skin tone and ground contact. The two eyelines must intersect in the same world.

Contact shadows, clothing occlusion and depth of field are what put people genuinely in one space. Without them, they read as stickers.

38.4 Hands and props#

Hands can be repaired or replaced independently while preserving wrist angle, skin tone, cuff and motion blur. Text on key props uses a tracked plane, with reflections, perspective and depth of field added.

If a prop rotates substantially and continuously, tracking cost is high — shot design should have it come briefly to rest before it is read.

38.5 Screen replacement#

Generate a clean screen with trackable corners and insert the UI in post. Handle perspective, brightness, black level, screen reflection, a sense of refresh, and finger occlusion. Screen light should fall on the character's face, or the content looks pasted on.

38.6 Weather and effects#

Rain, smoke, sparks and effects split into foreground and background and obey direction, speed and lighting. Effects illuminate characters and environment, so interactive light must be added. Particles must never obscure identity or proof objects.

38.7 Edges and motion#

Hair, translucent fabric, motion blur and defocused edges are where composites break first. A static matte flickers under motion, so work frame by frame or use a more stable source. Difficult hair can be made easier to key by designing the background and lighting for it.

38.8 Unifying color and texture#

Match exposure and white balance first, then contrast, saturation, black level, sharpness, noise and grain. Differences in skin texture between models cannot be solved by applying one LUT.

The composite master keeps sufficient bit depth and a subtitle-free version; the final unified grade happens later.

38.9 Repair priority#

Fix narrative blockers first: faces, hands, evidence text, occlusion. Then space and lighting. Then fine texture. An audience will not overlook a wrong name because a background glass is slightly different.

38.10 Design for compositability during shot design#

A compositable shot gives post something to hold. A camera orbiting quickly while hair swings widely, the fingers occlude a phone screen, and the screen reflects complex lighting is theoretically compositable, and the cost may exceed redesigning the shot. During shot design, deliberately arrange brief holds, clean planes, trackable corners, relatively stable occlusion and sufficient resolution.

Each layered shot answers: which elements must genuinely interact at the same time and which can be generated independently; who occludes whom at the boundary; which light affects which subject; whether motion blur belongs to the source or is added; where the focal plane sits; and whether a shorter usable interval exists if tracking fails. Unclear answers mean there is no composite plan yet.

38.11 Plates, clean frames and patch material#

Any complex shot with characters or foreground occlusion should generate or retain a background plate. Plates are used to patch character edges, remove wrong objects, extend holds, stabilize walls and build push-ins. Also keep text-free screens, unoccluded props, foreground character mattes, separate shadows and local material textures.

None of that looks like finished footage, and it is the most valuable insurance post has. With only a flattened video, an edge flickering at a character's shoulder leaves the compositor no clean background — only blurring, or regenerating the whole shot.

38.12 Color matching is not applying a LUT#

Two layers must first make physical sense. Fix the order: exposure and dynamic range, white balance and cast, black level, light direction, local light and shade, saturation, skin tone, sharpness, noise, motion blur, unified grain. A LUT handles part of the overall mapping. It cannot produce the contact shadow under a character's feet, and it cannot make a key light on the right fall on someone composited in later.

Character edges often show a bright fringe, either from spill in the original background or from insufficient matte contraction. Remove the original background color first, then rebuild the edge light of the current environment. Edges that are too hard read as cut paper; too soft and they glow. Review at normal playback speed and at 100 percent pixels.

38.13 Four levels of realism for screens and documents#

Correct graphic content is only the first level. The second is geometry: perspective, curvature, lens distortion and tracking. The third is imaging: brightness, black level, refresh, moiré, depth of field and motion blur. The fourth is interaction: finger occlusion, glass reflection, screen light on the face, a fold in paper pressing across text, ink shifting with the shadow on the page.

Design phone content at real screen dimensions first, then place it into the shot — do not typeset directly onto the video frame. Key evidence still needs its own readable insert; do not rely on the audience finishing three lines of a contract on a moving handheld screen. Graphic assets bind to the fact ledger too: names, amounts, dates and message senders are never written ad hoc by the compositor.

38.14 Three routes for two-person contact#

The first is single-source generation, suited to short, simple contact with acceptable identity risk. The second is layered pseudo-contact: foreground and background, occlusion, reaction and sound effects convince the audience that the handover, shove or catch happened, without holding on the actual point of contact. The third combines a locally photographed or purpose-shot action element with AI characters, for gripping, embracing and fighting.

Choose by whether the contact carries story evidence — not by which route is most purely AI. If the ring on the killer's hand must be seen as he grabs the protagonist, the contact point is a factual shot needing higher control. If it only has to convey that someone caught her from behind, shoulder occlusion, a sudden stop, fabric sound and a reaction shot will complete it.

38.15 Version structure for composite projects#

A composite master must be able to return to its components. Store at minimum the project version, the source material list, layer order, mattes, tracking data, graphic asset versions, color nodes, font licensing and output settings. The final filename cannot simply be shot_final.mov; it associates with a shot ID and a composite version.

When upstream assets change, use the dependency relationships to determine what must be recomputed. Changing the date in a phone message updates the graphic layer and re-outputs. Changing a character's cuff may affect the character source, mattes, contact shadows and grade. Without a componentized project, any small change degenerates into redoing the whole shot.

38.16 The phone livestream shot in Backlit Takeover#

After the banquet blackout, Lin Yun sees what appears to be her mother under duress on her phone. The team first tried generating Lin Yun, the phone, the person on the stream, scrolling comments, emergency lighting and a panicking crowd in one pass. Every element was partly there and none was credible. The final approach used six layers: the blacked-out banquet plate, Lin Yun's foreground performance, the phone and hand, the approved stream footage, screen reflection with interactive light on the face, and background crowd silhouettes.

The stream content was produced separately with its identity locked. The phone rotates only slightly and holds still for 0.8 seconds at the key frame. Comments do not scroll in the episode at all — only a "stream interrupted" state remains. The red emergency light falls on hand, face and table together. What the audience sees is one complete shot; in production it is six independently editable layers of fact.

38.17 SOP for compositing#

First, decide at shot design whether this is single-pass or layered. Second, design layers, perspective and occlusion. Third, generate clean backgrounds and subjects. Fourth, complete screens, props and effects. Fifth, add contact shadows and interactive light. Sixth, match depth of field, motion blur and color. Seventh, check edge flicker frame by frame. Eighth, output the master and record parent assets.

38.18 Fault tree#

Symptom: characters look like stickers. Contact shadows, shared lighting and depth of field are missing. Rebuild the spatial integration.

Symptom: screen text floats. Perspective or tracking is inaccurate and reflections and occlusion are absent. Re-track and composite interactive light.

Symptom: hair edges flicker. The matte is unstable. Improve the source background, reduce motion, or repair frame by frame.

Symptom: a large effect carries no force. There is no environmental response, interactive light or sound. Let the world react to the effect.

38.19 Checklist, exercises and deliverables#

Check that layers were planned in advance; that perspective and eyelines agree; that contact shadows exist; that screens have reflection and occlusion; that effects affect the environment; that edges are stable frame by frame; that color and texture are unified; and that provenance is traceable.

Exercise one: break the E006 phone shot into layers. Exercise two: list ten matching conditions for a two-person composite. Exercise three: build a text replacement plan for a document. Exercise four: judge whether one complex shot should be regenerated or composited.

Deliverables for this chapter: the composite plan, the layer manifest, tracking data, mattes and masks, graphic inserts, the color match, and composite QC.

38.20 Composite preflight: decide before generating which layers must be separate#

Complex shots tag background, characters, foreground occlusion, hands and props, screens, weather and light effects at the shot design stage, then judge which layers need independent control and which can be generated in one pass. Text, replaceable screens, high-risk hands and crossing character occlusion are usually worth separating.

Preflight also requires plates, clean frames, tracking marks and matte headroom. Discovering after generation that there is no clean background makes patching significantly more expensive. The composite plan enters the keyframe request rather than being improvised privately in post.

38.21 Layer ownership and spatial relationships#

The layer manifest records world depth, occlusion order, light sources, depth of field, motion, color space and source asset for each layer. Lin Wei's shoulder is foreground, Lin Yun is background, the authorization sits at the table's middle depth; when a hand crosses the document, the occlusion order changes over time. Without depth relationships, the composite reads as stickers.

Each layer has exactly one authoritative source. Screen text is controlled by the graphic registry, character identity by the character assets, shadows by compositing. Do not leave gibberish in the base image and stack new text on top of it. When one layer changes, the dependency graph invalidates only the results that need it.

38.22 Four viewings for composite acceptance#

The first pass at normal speed asks whether attention breaks. The second goes frame by frame for edges, giveaways and tracking. The third raises and lowers brightness to check black fringes, spill and mattes. The fourth uses a phone at the target bitrate to check fine lines, text and grain. Technically seamless still has to pass narrative legibility.

Common light, contact shadows, reflection, motion blur, grain and lens breathing together make layers belong to one camera. A global LUT cannot substitute for contact relationships. Passing results enter the composite manifest along with the project file, proxies and hashes.

A note on sources#

Critical text, multiple people and effects are frequently unsuited to single-pass generation. Treating compositing as a designed route protects narrative fact and reduces repeated rolls of the dice.