WARDROBE A→B · SEPTEMBER 5, 2026
Camotion endpoints and Cinematographer locomotion language appear complementary.
This chapter records a controlled Seedance 2.5 ablation on the Integration Test 01 A→B pair. It is not Camotion 01.13. The 01.8 operator was not changed. No images or Camotion frames were regenerated.
Notation: A / B = pristine canonicals. A′ / B′ = Camotion-conditioned shooting frames. A→B = generated traversal shot.
Camotion is motion-state conditioning, not image enhancement.
01.12 remains true on Ghost Library stills: there was no useful pristine-source operating window where substantial baked travel could be encoded without objectionable structured copies. Today's video evidence is complementary, not contradictory. A′ does not need to be an aesthetically superior still. It may function as machine-facing motion-state conditioning for the video model.
Three kinds of A→B evidence, do not mix them
PROVENANCE
Exploratory web UI — three manual Replicate generations, different seeds (50 / 40 / 60), web-UI aspect_ratio 16:9. The third used a shorter manual locomotion prompt that is not the Integration Test 01 prompt. Packaged under experiments/ab-exploratory-web-ui/.
Original Integration Test 01 A→B — conditioned A′/B′, authoritative CM locomotion prompt, seed omitted, previously judged the strongest “killer” traversal of the unattended loop.
Controlled 2×2 — the four videos below. Same seed 70, aspect_ratio adaptive, generate_audio false, sequential, no retries. Independent variables: pristine vs conditioned endpoints, and minimal journey vs the authoritative IT01 prompt.
CONTROLLED 2×2 · SEED 70
Minimal journey prompt, verbatim: A continuous journey from a cozy attic bedroom, through the old wardrobe, and into the snowy winter world beyond. Motion prompt: the exact Integration Test 01 A→B Seedance prompt from generation-manifest.json. COMMON_VIDEO_INTENT was not prepended.
01 · pristine + minimal
SEED 70
prediction g4ygbpr4qsrmy0d0ec891myw64 · A.png + B.png
02 · pristine + CM locomotion
SEED 70
prediction 50m7gg6nfdrmt0d0ecbrc8j618 · A.png + B.png
03 · Camotion + minimal
SEED 70
prediction d2151cn9tsrmw0d0ecfadxnew8 · A′ + B′
04 · Camotion + CM locomotion
CLEAR WINNER · ONE SEED
prediction 8kwb28tayhrmr0d0ecnaca4nzr · A′ + B′ · IT01 locomotion prompt
Independent human review: strongest spatial and temporal continuity of the four. Neither pristine+locomotion nor conditioned+minimal matched the combination. The wardrobe/clothes → snowy forest passage was judged even better than the original unseeded Integration Test 01 A→B. Bedroom → wardrobe threshold → hanging clothes passing the camera → dark intermediate volume → snow. No intermediate canonical was supplied.
Retain Camotion + Cinematographer prompt design in Phase 1.
This is one controlled seed, not statistical proof, not a claim that Camotion is universally necessary, and not a claim that the current operator is optimal. It is enough evidence to keep both complementary pieces in the working pipeline: the CM prompt supplies semantic/physical locomotion intent; Camotion supplies visual motion-state conditioning; Seedance combines them.
Spatial continuity is not temporal continuity
EASING
Forward motion is not the same as velocity continuing through B. Seedance often appears to ease/resolve toward the endpoint, so B can behave like a stop rather than a sampled position on a continuing trajectory. Canonical frames should be position samples, not stop points. 04 was perceived as better at retaining ongoing motion at B.
A crude exploratory frame-to-frame visual-change check performed outside the repo suggested increasing final-half-second activity 01≈0.96, 02≈3.47, 03≈4.32, 04≈5.33. Those numbers are not camera velocity. They can include snow, morphing, and exposure change. Informal observation only.
Do not mechanically insert intermediate canonicals
FUTURE CM
04 did not receive an extra canonical X. The video model invented useful intermediate wardrobe volume from conditioned endpoints, geometry, and detailed route language. Future CM reasoning: can this pair be covered as one continuous spatial move? If yes, describe/condition the route. If no — as in the Integration Test E→A flat door — recommend an intermediate canonical. Not implemented here.
EXPLORATORY WEB UI · DIFFERENT SEEDS
These motivated the 2×2. They are not controlled. Web UI aspect_ratio was 16:9. Run 1 generated audio. Run 3 used a shorter manual locomotion prompt, not the IT01 prompt. Informal Downloads filenames said “nomotion”; the prompt was the minimal journey sentence.
Pristine + minimal · seed 50
EXPLORATORY
prediction pzvvne7bc1rmt0d0ebdarwwx28 · generate_audio true · originally A-B nomotion.mp4
Forward motion occurred. Transition felt like a cut/crossfade/dissolve. Easing toward the endpoint.
Camotion + minimal · seed 40
EXPLORATORY
prediction fh8t4h3r7xrmy0d0ebetfpn6am · originally A'-B' nomotion.mp4
Forward motion; somewhat less cut-like than pristine/minimal, still a cut/stop rather than fully continuous traversal. Easing remained.
Pristine + manual locomotion · seed 60
EXPLORATORY
prediction ey3nfwjrqdrmt0d0ebssagmr2c · originally A-B motion.mp4 · prompt is not IT01
Forward motion. World transition still felt like a cut/dissolve. Do not rewrite history: this is not the Integration Test 01 prompt.
After this checkpoint
The remaining bounded Camotion question was whether the Cinematographer can choose a small scene-appropriate 01.8 strength from an already safe range. That experiment is now complete: scene-aware {.02,.04,.08} was preferred over fixed .08 on three Wardrobe Loop pairs. Camotion Phase 1 is frozen. The later shootability / intermediate-volume experiment is also complete: see chapter 09 and the 2×2 report.