Creativity + editable workflow

Why MetaHumans. Every Unreal render becomes a reference video for the generator. Stand-in characters with simple textures and linear animation get copied: broken joints, cartoon surfaces, stiff motion. MetaHumans already look close to the final cast, in the final clothes and hair, so the reference teaches the right thing.

One set, every angle. Eyelines, screen direction and where each person stands stay consistent across close-ups and reverses, because every shot is filmed in the same 3D space.

Why not the plain render? Fed the full-color render, the generator copies the MetaHumans' faces and their animation too closely. Fed only a depth map, it reinvents the set in every shot. The Depth Ghost splits the job: color pixels hold the location, gray depth holds the staging, and the performance is generated fresh.

Built in ComfyUI. One reusable ComfyUI graph makes every Depth Ghost, laid out node by node above. It masks the actor with SAM3, feathers the mask, estimates soft MiDaS depth and composites that depth over the full-color render, so only the person turns gray.

Soft on purpose. In the ComfyUI graph, MiDaS estimates depth at 384 pixels and the result is scaled back up, so the silhouettes have fuzzy edges that blend into the set instead of reading as cutouts. Custom Python scripts steady the brightness across each shot so the guide doesn't flicker.

What it taught me. Those soft edges also gave the generator room to repaint what sat right behind the characters, so the doorway drifted between shots. The fixes are in step 04.

Restraint has to be written down. Generated actors default to big, fast emotion. Prompts spell out a slow build, which muscles carry the feeling (brows, lids, mouth corners, chin, breath), eyes that stay dry until the right moment, and bodies that stay alive while standing still. A real voice recording guided the emotion and pace of the delivery.

Recolor by color key. The early door shots drifted to a cool gray interior and a red door. A per-shot color key isolated just those areas and moved them to the film's warm brown, with edge cleanup around her hair. No new generation, no roto.

Replace the background. The tears-and-sleeve-wipe take was the best performance but had the wrong house behind him. The actor was matted on every frame and placed over a generated clean plate of the correct doorway, with Claire breathing in soft focus, lens-style blur, edge color cleanup and a subtle handheld drift so both layers share one camera.

The problem. Topaz Starlight redraws the finest lines (lashes, single hairs, iris rings) razor sharp while smoothing the texture in between (pores, skin sheen, fabric). The result is detailed but waxy: realistic, yet not natural.

The fix. Frequency separation. Color and real texture come from a plain upscale of the original footage; only the finest lines come from Starlight, at a strength set per shot and turned down on face close-ups. Fine grain ties it together. One command per film.

Unreal established the geography. A roughly 2.3 km wave and seven-ship fleet supplied the scale cues. Reusable ship geometry supported different camera positions.

The final image developed the look. Cloud cover, foam, spray and surface detail were directed through reference images and AI generation. The Unreal ocean served previsualization, rather than a finished fluid simulation.

Reusable geometry, adjustable coverage. The LPD was prepared from a supplied model, scaled and brought into Unreal. A separate open-stern variant supported the boat-launch setup. The orbit could be revised without rebuilding the ship for each angle. Deck, hangar and aircraft remain concrete landmarks.

The useful constraint was contact. The stand-ins established who occupied the boat, where Amy stepped off and how the camera followed the ascent. The final shot replaced proxy anatomy and water materials while retaining the story action.

Build for the lens. The exterior aircraft and interior filming set were separate. The selected camera and its preview show how the interior coverage was composed inside a simple box suspended beside the wave.

Unreal controlled the action arc. Separate rotor components, aircraft transforms and camera keys established the exterior movement. Cloud occlusion gave the descent a clear visual destination.

The final supplied atmosphere and detail. The finished shot adds realistic aircraft surfaces, dense clouds and darkness. It follows the descent beat; the framing and timing are not a frame-for-frame render of the animatic.

Sequencer held the scene together. The summit assembly moves from arrival and Mother's interruption to refusal and singing. Stand-in motion communicates intent; final human movement and expressions are directed separately.

Unreal was the place to inspect timing. The leap used authored skeletal animation and a ballistic trajectory. Unreal provided a place to inspect timing, distance and camera proximity before the final creature and performance were generated.

The depth video is a guide, not a performance lock. The production required iteration when generated shots copied stiff poses or froze the water.