Method, evidence, and limits
Every result depends on a renderer, source field, preprocessing choice, metric implementation, resolution, sample interval, control space, and declared intention. This page separates what was measured from what was inferred.
Plain-language system map
ISR and VTL do different jobs. ISR turns a rendered frame into inspectable visual evidence. VTL measures relationships within that evidence. A controller then changes only the scene inputs it has been authorized to change.
ISR, the Image Structure Reader, derives fields from perceptual lightness, edges, slower tonal variation, chromatic activity, and gaps. Thresholds and masks turn those fields into regions that can be measured.
In ordinary language: it separates where visual activity is happening from where it is quiet, without first deciding that a region is a person, doorway, or story event.
VTL measures placement, void, packing, cohesion, boundary engagement, orientation, structural thickness, dispersion, G1–G4 behavior, color relationships, and selected composites. Across frames, it follows how those values travel, fluctuate, or settle.
In ordinary language: it can report that the frame’s structural center moved, that edges became more engaged, or that the organization changed less through time. It does not automatically know why that matters dramatically.
The original note is retained, then expressed as targets to pursue, protections that must remain within tolerance, observations to disclose, and free channels that may move.
This translation is a hypothesis, not a replacement for the note. The director judges whether the measured result actually fulfills the intended relationship.
The controller varies camera, lens, aim, blocking, light, material, or other authorized inputs. Blender or Unreal renders the consequence. Invalid candidates are rejected; valid alternatives guide the next local search.
The result remains an editable scene decision with its inputs and measurements attached—not a score pasted onto a finished movie.
One frame, traced end to end
This is frame 24 from the retained five-character Unreal baseline. Every panel comes from that exact render using the implemented ISR lightness and edge fields plus the VTL structural mask. The plate is diagnostic output, not an artist’s reconstruction.
The structural center sits 2.78% of frame width right of center.
It sits 1.46% of frame height above center.
Active structure occupies a relatively sparse share of its convex hull.
Most selected pixels belong to one connected structural component.
4.26% of masked edge weight falls inside the outer 8% band.
Edge directions are broadly distributed rather than strongly concentrated.
The selected 2D structure is thin relative to the frame size.
Masked mass has this normalized spread around its own centroid.
What it reveals
The center is only slightly right and high, with little structural weight at the frame boundary. ISR retains the figures and the main architectural relationships as one inspectable surface.
What it does not know
The high cohesion value does not mean “the family is united.” The wall, floor boundary, doorway, and figures form a largely connected mask. VTL reports that visual fact; character identity and dramatic meaning require scene data and interpretation.
How time enters
The same read is repeated across the shot. The temporal layer then compares paths, ranges, reversals, crossings, and total variation rather than pretending this one frame contains the whole beat.
What each layer can actually know
| Layer | What it reads | Useful translation | Boundary |
|---|---|---|---|
| ISR evidence fields | Lightness, edge activity, broad tone, chromatic activity, gaps, masks, and connected regions | Where the rendered surface is active, quiet, continuous, fragmented, bright, or chromatically distinct | Regions are not automatically named characters, props, emotions, or narrative functions |
| VTL spatial coordinates | Centroids, void proportion, packing, cohesion, edge-band load, orientation, thickness, and dispersion | Where structure sits, how tightly it gathers, how it engages boundaries, and how stable its organization appears | A high or low value is not inherently better; the brief supplies the desired direction |
| VTL temporal coordinates | Per-frame paths, ranges, total variation, crossings, reversals, and changes across a sequence | Whether a visual relationship relocates, oscillates, accumulates, or remains settled while visible action continues | Sampling is not eye tracking, and structural change is not automatically dramatic meaning |
| Scene and geometry checks | Actor roots and bones, projected figure scale, landmark visibility, camera state, animation, and scene identifiers | Whether the doorway, witnesses, performance, scale, screen direction, or other production facts were actually preserved | These protections come from Unreal or explicit validation—not from VTL alone |
Worked translation
“Keep the doorway and all three witnesses. Make the arrival split the group more decisively, but do not make the frame frantic.”
This is a hard protection. Reject any candidate that crops the landmark, loses a required body, or breaks the requested geography. Evidence comes from projected scene geometry and visibility checks.
This is the target. Search for a stronger change in group separation and temporal placement around the arrival, then inspect whether the render reads as the intended fracture.
This is a restraint bundle. Bound edge pressure, center reversals, and unwanted temporal volatility rather than minimizing one universal score.
This is causal protection. Lock animation and actor transforms, then verify them across every sampled frame so the camera cannot borrow its result from changed acting.
The essential distinction
VTL does not “understand” the directing note by itself. It makes a proposed translation testable: the note remains primary, measurements reveal consequences, scene checks enforce facts, and human review decides whether the visual choice earns the language.
Closed-loop architecture
Desired relationship, protections, freedoms
Camera, lens, aim, scene controls
Perspective, bodies, architecture, light, void
Coordinates, reliability, temporal paths
Targets and preservation gates
A kernel does not directly dolly or rotate a camera. The controller changes authorized inputs, the renderer produces the consequence, VTL measures that consequence, and the note determines which candidates remain eligible and where the next search begins.
Evidence discipline
No-change is an alternative. Difference is not rewarded merely for being different.
Unreal trials compare roots and named bones over every sampled frame, preventing camera results from borrowing altered acting.
A coordinate may be pushed while scale, landmarks, visibility, or another channel stays within tolerance.
“Δx range increased” is a measurement. “The entrance feels disruptive” is a proposed reading requiring human validation.
Repeat renders establish the ordinary variation of each coordinate. A claimed effect must clear its own observed stability envelope.
Focused production calibration
This is a local stability check for the public Sprite Fright example—not a universal calibration of VTL. It uses the same renderer, shot, resolution, evidence substrate, and measurement code as the joint-directing result.
Frames 103, 128, and 154 were rendered independently along the camera-only and joint routes before the blocking intervention began. Their parameters were identical.
At frame 256, the camera and performance were held while the lead rig stepped through 0.00, −0.25, −0.50, and −0.75 m on one world axis.
An adjacent step clears the local floor when its change exceeds three times the largest observed repeat difference. That supports detection only—not perceptual or dramatic value.
| Channel | 0.00 m | −0.25 m | −0.50 m | −0.75 m | Local result |
|---|---|---|---|---|---|
| Δx · horizontal placement | −.0398 | −.0443 | −.0548 | −.0581 | Monotonic; every step clears floor |
| ρᵣ · packing | .4496 | .4618 | .4722 | .4739 | Monotonic; response compresses at far end |
| μ · cohesion | 1.000 | 1.000 | 1.000 | 1.000 | Saturated; unavailable for this local choice |
| xₚ · peripheral pull | .0249 | .0252 | .0261 | .0264 | Small but above observed floor |
| SDI · dispersion | .1550 | .1521 | .1464 | .1446 | Monotonic; every step clears floor |
| G1 path length | .0640 | .0663 | .0695 | .0709 | Monotonic; every step clears floor |
| G1 reliable curvature | .1559 | .1005 | .0822 | .0463 | Every step far exceeds .000035 repeat maximum |
Useful: placement, packing, dispersion, and both G1 measures respond consistently to the blocking axis in this frame.
Diagnostic: cohesion is pinned at its maximum and cannot distinguish the alternatives. Structural thickness stops moving at the largest step.
Boundary: three repeat pairs and one scene axis do not establish a general noise distribution, a just-noticeable visual difference, or stable behavior in another renderer.
Claim ledger
The table is deliberately conservative. A narrow claim that survives inspection is more useful than a grand claim assembled from suggestive labels.
Exact renders, inputs, masks, coordinate versions, and selected outputs can be retained together.
Coordinate trajectories distinguish structural behavior from raw frame difference and visible movement.
Rendered feedback can find inputs that move a requested relationship while explicit protections remain within tolerance.
The same note was realized through camera-led, blocking-led, and balanced routes while performance and required geography remained protected.
Three notes accumulated across one 16-second Unreal sequence while approved time ranges and untouched character and camera channels remained locked.
A five-character Sprite Fright shot accepted two cumulative camera notes and one selective blocking note while the original source remained untouched.
No current test establishes that “suspense,” “release,” “disruption,” or “inevitability” transfers reliably across scenes or audiences.
The instrument does not infer whether an image was deliberately made, nor does it possess a validated independent aesthetic preference.
Camera and scene edits change a single projected surface. Coordinates couple, sometimes nonlinearly, and must be protected explicitly.
The current notes were encoded as experiment logic. A director-facing system that translates new language into safe, editable revisions has not yet been built.
Scale-space paths and sampled frames are measurements of representations; they are not eye tracking or direct cognitive-duration measures.
Failures that improved the system
An engine conversion left 85.3% of pixels below 5% luminance. The resulting narrow classification was a render-validity problem, not a discovery about the scene. Preserving Cycles corrected it.
Lesson: renderer and illumination belong to the evidence contract.Peripheral and directional briefs reached their targets but converged visually because diagonal structure and figure placement were not protected.
Lesson: “leave the others alone” must become named tolerances.The first branch reduced change partly by weakening projected figure scale, even though packing density looked acceptable.
Lesson: a plausible metric proxy does not replace the visual property actually requested.The disruptive shot had lower peak structural G1 than the inevitable shot while reorganizing more over time.
Lesson: within-frame scale-space stability and temporal reorganization are separate coordinates.State-identical protected frames varied by only .067 of one 8-bit pixel level on average, yet threshold-based ρᵣ and μ moved as much as .0404 and .0437. The same audit found Δx stable to .0010.
Lesson: every coordinate needs its own observed render floor before a small effect is interpreted.Reproducibility
Scene files, animation assets, exact frame ranges, and performance specifications.
Camera position, aim, lens, scene parameters, random seeds, and search domain.
Frame sequences, review movies, engine settings, resolution, and hashes where recorded.
Raw coordinates, reliability and availability, temporal summaries, selection records, and warnings.
Native Unreal Level Sequences with animation, transform, and camera-cut audits.
Original directing language, provisional coordinate translation, visual review, corrections, and claim boundaries.
Credits and version provenance
Attribution, software version, measurement layer, disabled systems, implementation hashes, and claim boundaries travel with the public artifact rather than depending on the development computer.
© Blender Foundation. Attribution: (CC) Blender Foundation | studio.blender.org. Project material is licensed under Creative Commons Attribution 1.0 except where specifically indicated otherwise.
Official license and attributionSequences: 960 × 402 at 32 samples. Matched hero frames: 2048 × 858 at 160 samples. Source files were never saved over; embedded production scripts remained disabled.
The foreground RGBA render is authoritative for attributing camera and character changes. The presentation composite checks the visible whole frame. Scene geometry separately verifies named actors and clipping.
Kernel and temporal-source hashes are recorded because historical notebooks with similar labels are not assumed numerically interchangeable. The public record contains no local machine paths.