The still-image sequence was frozen before motion testing. The video experiments then treated selected stills as spatial anchors and asked a video model to infer the unseen traversal between them. The key distinction was not visual smoothness, but whether the camera appeared to move through a stable world instead of morphing, dissolving, or cutting from one endpoint to the next.
Model: Nano Banana Pro
Format: 16:9 · 2752 × 1536
Model: Seedance 2.5
Baseline: 720p · 6 seconds · first/last frame only · no @ references
The wrong exterior frame was accidentally used as the endpoint. The result nevertheless produced convincing physical traversal from the cave into an exterior environment. Because the endpoint was not canonical B, this clip is excluded from canonical sequence evaluation, but it established that the method could work at all.
Seedance physically approached the cave opening, crossed the threshold, revealed the exterior geography, and arrived convincingly at B. This was the strongest evidence that a video model could infer unseen connective space between two independently generated stills.
Important imperfection: after exiting the cave, the terrain implies a step down, but the camera remains at a nearly fixed vertical height. The result feels more like a stabilized drone or camera rail than an embodied person walking. Spatial continuity passed; terrain-responsive body motion was only partial.
The clip remained visually smooth, but the world reorganized around the camera. Roots and earthen banks gradually reconciled into the destination instead of being passed in stable world space. Optical forward motion was present, but the dominant mechanism was a slow morph/dissolve rather than physical traversal.
No embedded clip was retained for this attempt. The observed failure changed from a gradual morph to an abrupt hard cut: stronger language suppressed the dissolve behavior without producing the missing physical route. Prompting harder did not solve the geometry problem.
X₁ was generated with Nano Banana Pro using B and C as references, explicitly asking for a plausible intermediate camera position between the two. It is a traversal waypoint, not a canonical story frame.
The shorter spatial jump still did not induce genuine locomotion. Seedance largely held the camera position and slowly reconciled B into X₁, suggesting that waypoint insertion alone was not sufficient.
Reducing duration did not materially rescue the traversal. This controlled test weakened the hypothesis that the 6-second duration was simply too long for the apparent distance between B and X₁.
This test deliberately asked whether the strong A → B result would generalize to another threshold-like transition later in the journey. It did not. The clip remained visually coherent, but the camera made little meaningful forward progress; the root, ledge, and mushroom structures reconciled toward G around the viewpoint instead of being physically traversed.
Interpretation: a narrative threshold is not enough. The endpoint images themselves must contain compatible spatial evidence that allows the video model to infer a route through a stable environment.
Why we tried it: The reel suggested B → C barely encoded camera movement. Skipping C created a larger visual jump and tested whether a more clearly different endpoint would encourage actual traversal.
Result: It still favored dissolve/reconciliation over locomotion. A larger endpoint difference alone was not enough. One earlier attempt at this same test was invalidated by Seedance-generated audio triggering a copyright check; that run was excluded from the visual traversal evaluation.
Controlled comparison: The anchors were unchanged from Test 07. Only the prompting philosophy changed.
Result: Dramatically stronger physical travel. Foreground trunks and roots sweep past the camera, occlusions happen naturally, new terrain is continuously revealed, and strong parallax makes the shot read as movement through a space rather than B dissolving into D.
Important detail: Large foreground roots and trunks function almost like natural wipes. They give the model opportunities to reconstruct unseen geometry while preserving the viewer's perception of uninterrupted travel. Incidental motion such as the falling leaf adds temporal texture and additional optical-flow evidence without being the core reason the transition succeeds.
Remaining limitation: The traversal succeeds perceptually, but the clip still eases in at the beginning and eases out at the end. That is acceptable for a single shot, but repeated easing would expose every clip boundary in a longer stitched journey.
In an August 29 check-in, Terran Boylan explained an important part of his original TunnelVision workflow, later used in Digging in the Dirt, that had not yet been represented in this log: before video generation, he preprocessed the keyframes in Photoshop so they visually communicated that the camera was already moving. His script used two different blur operations, incorporated Z/depth information, and protected the intended destination area from blur.
The practical effect is a depth-aware motion cue: nearby structures can streak strongly while the destination remains comparatively stable. This means the original technique combined linguistic motion conditioning in the prompt with visual motion conditioning in the endpoint frames.
Credit boundary: motion-conditioned keyframes, depth-aware blur, destination protection, and the generic continuous-motion prompting strategy come from Terran's original TunnelVision work. The experiments below test extensions of that approach inside the emerging agent architecture.
Result: Convincing physical traversal from the old library through the doorway into the courtyard. Shelves, skeletons, and foreground structures pass the camera with strong parallax; the doorway functions as a real threshold; and the exterior is revealed after crossing it rather than dissolving into place.
Accidental control: The prompt still contained the earthen-ravine language from the previous forest experiment. Despite the semantic mismatch, the visual anchors dominated and the traversal remained strong. This suggests that the motion-conditioned endpoints and continuous-locomotion grammar are doing substantially more work than elaborate scene-specific prose.
New extension under test: Terran did not use pristine canonical images as additional references in his original TunnelVision workflow. Here, the pristine library frame was supplied as a visual reference while the motion-conditioned first/last frames remained the motion authority.
Result: Also strong, but not dramatically better than Test 09. The scene-aware language correctly describes the library, doorway, skeletons, courtyard, parallax, and ghostly atmospheric motion, yet the overall traversal behavior is broadly similar.
Interpretation: The Cinematographer probably should not ask an LLM to freely rewrite the whole video prompt for every shot. A stable, proven locomotion template can be filled from structured scene data: environment, route, destination, foreground occluders, atmospheric motion, speed, and slight steering. The more valuable reasoning effort may be geometric: where to go, what to protect, what should pass the camera, and how to condition the anchors.
After Tests 09–10 we distinguished a pristine canonical frame from a derived motion-conditioned frame. At this stage we treated the canonical image as authoritative world state, and hypothesized that pristine originals might also be supplied to the video model as visual/reference authority for geometry, materials, lighting, and fine detail, while the conditioned frames remained the motion authority.
We noted that a controlled test should compare clean S/E, motion-conditioned S/E, and motion-conditioned S/E plus pristine references. The library experiment at this stage had only the pristine start frame available. Later controls, recorded in the Camotion checkpoint below, changed this: extra pristine references are not current architecture, and they did not explain the recursive-library artifact.
Visual continuity is not spatial traversability, and endpoint matching is not the same objective as locomotion. Our destination-oriented prompts repeatedly encouraged Seedance to reconcile one frame into another. The successful B → D controlled test used the same anchors but changed the instruction hierarchy: continuous movement became the governing fact of the shot, while the environment was allowed to negotiate bends, occlusions, particles, and foreground objects in service of uninterrupted travel.
This confirms the value of Terran Boylan's original TunnelVision motion language rather than replacing it. The new opportunity is to generalize that principle inside a Cinematographer agent: inspect the current world, infer appropriate route-taking, occluders, particles, moving objects, and motion cues for that environment, then construct a world-specific traversal prompt whose first priority is never stop moving.
This was not only an experiment in spatial continuity. At every iteration we were also making directorial and aesthetic decisions. Those selections changed the reference image for the next generation, which meant preference could propagate visually even when it was not restated in the text prompt.
The first exterior attempt still showed the cave. We then explicitly placed the tunnel behind the camera. Later, literal distance became less useful than asking for the next meaningful camera position and requiring the previous foreground to be left behind.
We consistently preferred images that were ominous, uncanny, strongly framed, and increasingly strange while still feeling physically connected to the same environment. Organic root architecture, enclosed routes, thresholds, dramatic scale changes, and unresolved destinations repeatedly won over safer or merely attractive alternatives.
The early prompts explicitly described geology, light, roots, soil, camera position, and visual continuity. As the world became established, the reference image carried more of that information. The text prompt could increasingly focus on movement, continuity, and what route to follow rather than repeatedly art-directing the image.
Choosing an image made its visual characteristics part of the next generation's context. We did not need to keep saying “make it more ominous” or “use larger roots.” By selecting the versions that embodied those qualities, the aesthetic preference traveled forward through the reference image itself.
The glowing mushrooms were not planned. They first appeared as an incidental visual discovery. We selected the image containing them, they persisted in the evolving world, and eventually we explicitly followed them in the navigation prompt. A generated aesthetic detail became a functional route through the environment.
The strongest images did not introduce arbitrary fantasy objects. The world became stranger by exaggerating things already present: exposed roots became architecture, an opening became a threshold, a chasm changed the scale, and a giant root became a crossing. Novelty felt discovered rather than imposed.
It must preserve enough visual and geographic continuity to feel like the same journey, while introducing enough new information to change what the viewer wants to know next. This became our working directorial test for candidate selection.
Every selection and rejection contains information about both navigation and taste: forward progress, curiosity, atmosphere, composition, scale, thresholds, and meaningful environmental discovery. A TunnelVision agent could accumulate those signals and use them when proposing future routes and evaluating candidates.
The lightbox reel made B → C, D → E, and F → G look much weaker spatially than they had during pairwise selection. The sequence could be aesthetically progressive while still failing to communicate that the camera physically changed position.
A Director should not evaluate only the current candidate against the immediately previous frame. It can maintain the canonical reel, periodically audit the trajectory, detect near-stationary viewpoint changes, and regenerate weak links before motion generation begins.
With identical B and D anchors, the baseline destination-oriented prompt dissolved between scenes while the Terran-derived prompt produced convincing travel. The successful strategy emphasized constant forward motion and let world-appropriate occlusions and foreground events help bridge imperfect geometry.
The successful B → D traversal still eases in and out. For a single clip that feels polished; across many clips it would create a repeated accelerate → travel → decelerate rhythm that exposes every stitch.
The ideal A → B | B → C stitch, as we then understood it, preserves the exact same canonical B on both sides while maintaining through-motion. This keeps the sequence visually exact without requiring trims that replace the shared anchor with two merely similar interior frames. Later we separated storyboard canonical identity from the shooting frames actually sent to video; B_in / B_out remains an open question.
Terran's library anchors produced convincing traversal even when the text prompt accidentally described a forest ravine. The visual motion signal and continuous-locomotion grammar appear to dominate much of the scene-specific wording.
Our first extension test used the motion-conditioned first/last frames to communicate movement while a pristine start image supplied clean visual identity. At the time this suggested a useful separation between canonical world truth and derived cinematography assets. A later control, recorded below, showed that removing the extra pristine reference did not eliminate the recursive-library artifact, so that extra reference is not treated as current architecture.
Scene-aware prose did not substantially outperform the mismatched generic prompt. The Cinematographer can instead output a structured shot specification that feeds both a stable prompt template and deterministic motion-conditioning code.
The Director can propose the entire canonical journey, while the user can intervene only where desired: open alternatives for a storyboard location, replace the Director's selection, or point to a different destination in the frame. Human selection is a source of direction and preference data, not a requirement for every step.
Earthen realism → eroded forest ravine → increasingly architectural roots → ominous threshold → impossible chasm → glowing mushrooms → root crossing → monumental tree opening.
The final aesthetic was never specified as a destination. It emerged through repeated generation, selection, and reuse of the selected image as the next reference.
Describe the desired scene in detail → force the old location behind the camera → test literal distance → request the next meaningful camera position → discover that selected frames can still leave the viewpoint nearly stationary → audit the growing reel for perceptible displacement → separate Director destination logic from Cinematographer locomotion logic.
The long-term agent opportunity is to automate both layers: the Director chooses a genuinely new, traversable camera position, while the Cinematographer inspects that world and generates an environment-specific continuous-motion strategy.
Chosen images carry aesthetic and structural preferences forward through the reference itself. Repeated choices become useful preference data even when those qualities are not explicitly written into every prompt.
The glowing mushrooms emerged unexpectedly, were selected, persisted, and eventually became an explicit navigational breadcrumb. Generated details can move from accident → preference → route.
A destination can belong beautifully to the same imagined world and still fail as a physical video transition. The Director eventually needs to reason about both desirability and reachability.
Give TunnelVision one current frame. The Director proposes several meaningfully different next-camera moves; the image generator renders them; an evaluator ranks continuity, perceptible camera displacement, novelty, navigability, preference fit, and discovery; the human chooses one; that choice updates the preference state and becomes the next canonical frame. The growing reel is periodically re-audited so aesthetically strong but spatially stationary transitions can be repaired before video generation.
Scope decision: automated video repair does not need to block the first working agent. Motion can be attached afterward as a Cinematographer/evaluator stage that attempts traversal, scores the result, and either accepts it or feeds failure information back into frame selection/regeneration.
The primary product surface is now a clean left-to-right Storyboard containing only canonical frames. The Director may generate the whole proposed journey from a starting image, story idea, and target duration. Clicking a storyboard location reveals candidate variations; selecting another candidate replaces that canonical frame and causes downstream frames to be re-evaluated. Clicking through canonical frames forward/backward remains the simplest human test of whether the journey actually travels.
A Direction view overlays the Director's chosen destination on a frame. The user can drag/click a different destination to override it. That normalized spatial intent becomes shared data for scene reasoning, route selection, vanishing-point/depth analysis, protected regions, motion conditioning, and prompt construction.