Direct in previs
Establish composition, action and camera timing. The technical colours identify objects; they do not define the finished look.
Case study 01 / Independent research
Previs-to-AI
Video Pipeline.
Can the image change
without losing the direction?
SourceMaya previs
Sequence4 seconds
Previs camera35mm
My rolePrevis · Visual direction · Evaluation
01 / The question
Preserve the action.
Change the image.
A frightened woman runs through a room and passes under a descending gate. The previs establishes framing, camera movement, character action and gate timing.
I tested whether those decisions could carry into AI video while changing lighting, materials, style, effects and the character. The goal was continuity—not simply a convincing still image.
This is an independent study, not a commissioned campaign. Even successful takes retain differences in pose and fine geometry.
02 / The process
One clear role for each input.
Define the appearance
Develop opening and closing frames for character identity, lighting, palette and materials.
Generate. Compare. Refine.
Use the video as the primary motion reference. Review the full sequence, then refine ambiguous or conflicting inputs.
03 / Selected experiments
Same shot.
Different possibilities.
Cinematic realism
Cold graphite materials and natural character detail. Strong staging correspondence, with small geometric differences.
A new visual language
A reference-led anime treatment, developed with Your Name references. The scene becomes a decorated castle: broad staging carries over, but the environment is deliberately reinterpreted.
Laser effects
Technical markers become orange-red light pulses. The broad action remains readable; exact shot-by-shot transfer and impact behaviour still require review.
Painterly stylization
An Arcane-inspired direction brings a painted treatment to the scene. This take remains closer to painterly realism than fully graphic animation.
Character replacement
A male character replaces the heroine in the cold, realistic environment. The broad action sequence survives; individual poses are not identical to the previs.
The effect starts in previs
Green cylinders specify the shots. Their technical colour is not the intended final colour. Pulses originate behind the heroine and travel in her direction of movement.
Visual development: the anime opening and closing frames
These developed frames establish the intended anime appearance. The castle treatment adds architectural decoration; it is not a claim of identical environmental detail.
04 / Where the shot broke
Useful failures.
Specific decisions.
Competing motion instructions
Observed: an early turn and changes in camera framing and action.
Response: return to a reference-led prompt without a second-by-second rewrite of the motion. The comparison supports this approach, but does not isolate prompting as the only cause.
Too much left to interpretation
Observed: the shot starts from behind; the environment is reinterpreted and the gate action loses its intended context.
Response: supply clear visual anchors and make the mannequin’s facing direction readable. Not every video-only test failed in the same way.
Material quality is a sequence-level problem.
Refining faces and materials in still images helped define the target, but generated footage could reintroduce noisy surfaces and unstable detail. A clean start frame is an input improvement, not a guarantee of clean textures throughout the shot.
05 / Findings
Control needs
critical review.
Visual anchors matter.
In these tests, image-led setups made the style more predictable than text-only direction.
Make the facing direction clear.
A faceless mannequin leaves front-versus-back orientation ambiguous. A readable face or directional marker can help clarify it.
Keep motion instructions consistent.
The simpler reference-led prompt worked better in our sequence than competing, detailed timing instructions. This is an observation, not a proven universal rule.
Watch beyond the opening frame.
Gate timing, body proportions, material stability and invented background details still need checking.
What these tests do not establish
They do not prove exact geometric reproduction, a universal success rate or guaranteed savings. Successful Recreate runs do not, by themselves, establish that the service fixes the random seed.
06 / Working configuration
The reference
comes first.
One video reference for staging. Two image references for appearance. A four-second generation, followed by comparison against the original shot.
This is the archived realistic baseline used in the study—not a universal template. “Exactly” describes the requested behaviour, not a measured result.
Read the working prompt
Use @[Video 1](video_1) as the primary source of truth for the entire shot. Reproduce its single continuous action exactly, including the framing, approved 35mm perspective, camera path, camera timing, character body motion, character scale, architecture, object placement, gate layout, gate position and gate animation. Do not reinterpret or improve the staging. Use @[Image 1](image_1) as the exact opening-frame appearance and @[Image 2](image_2) as the exact last-frame appearance. Use these two images only for the woman's identity, fuller body type, ordinary modern clothing, hair, lighting, color palette, materials and atmosphere. Do not use their geometry to override @[Video 1](video_1). The frightened fuller-built young woman has loose wavy blonde/light-brown hair and wears a cream knitted cardigan, a dark floral tunic or dress, blue jeans, white sneakers and a tan crossbody bag. Preserve her identity, body shape and clothing throughout. Reproduce her exact awkward running motion and timing from @[Video 1](video_1) without reinterpretation. Preserve the exact room architecture and gate geometry from @[Video 1](video_1). The interior has the same clean, monumental Harkonnen-inspired material language established most clearly in @[Image 2](image_2): heavy smooth blackened satin-steel gate bars and crossbeams, large quiet matte graphite architectural panels, and a dry honed dark-basalt floor with soft broad reflections. Cold graphite and steel-gray palette, restrained cyan-white practical lighting, deep detailed shadows and natural highlight roll-off. Keep every material temporally stable and physically coherent as the camera moves: no crawling patterns, texture swimming, flickering roughness, random grunge, crystalline or cloudy metal, wet patches or procedural noise. Keep natural skin and ordinary clothing readable. One continuous shot. No cuts, transitions or montage. No new camera movement, zoom, lens change, reframing, slow motion or handheld shake beyond @[Video 1](video_1). No new architecture, objects or people. No sci-fi costume. No identity, body-shape or gate redesign. No text, subtitles or watermark.
Tools: Autodesk Maya for previs; image generation and refinement; Seedance through Higgsfield for video; side-by-side review. Film and animation names identify visual references, not affiliation or endorsement.
The takeaway
AI can extend a directed shot.
Direction still means deciding what to keep, what to correct and when to try again.
The strongest results preserved the broad staging while allowing substantial changes in appearance. Fine geometry, body proportions, material stability and repeatability remained areas for supervision. These are findings from one sequence, not a benchmark of the model.