# Kling continuity test matrix Status: planned experiment matrix Captured: 2026-03-29 Purpose: systematically test which workflow patterns improve continuity, subject stability, and shot-boundary smoothness. ## Why this exists Our live tests proved that multi-shot works, but they did not yet reproduce the continuity quality implied by Kling's strongest reference materials. This matrix is designed to answer one practical question: **What combination of shot design + reference strength gives the best continuity for our pipeline?** --- ## Evaluation rubric Each test should be scored 1–5 on these dimensions: ### 1. Identity continuity - Does the same subject remain recognizably the same across all shots? ### 2. Scene continuity - Does the environment feel like the same place/time/lighting context? ### 3. Camera continuity - Do camera transitions feel like a natural progression rather than a reset? ### 4. Boundary visibility - How noticeable is the shot cut/jump at the transition points? - Higher score = less visible / less disruptive ### 5. Overall production usability - Could this output be shipped or edited into a real pipeline without embarrassment? ### Suggested summary metric `continuity_score = (identity + scene + camera + boundary + usability) / 5` --- ## Shared test principles These apply unless a row explicitly tests the opposite. - same environment - same subject - same time of day - same direction of motion - same overall emotional tone - adjacent shots should have high semantic overlap - no unnecessary angle reversals --- ## Phase 1 — shot-delta sensitivity Goal: find how much shot-to-shot change the model can tolerate before continuity breaks. ### Test C1 — minimal shot delta - Workflow: text-only multi-shot - Scene: same beach, same subject, same side-follow motion - Shot progression: - shot 1 = medium side-follow - shot 2 = slightly closer side-follow - shot 3 = slightly closer side-follow - Purpose: baseline best-case continuity attempt ### Test C2 — moderate shot delta - Workflow: text-only multi-shot - Scene: same beach, same subject - Shot progression: - shot 1 = medium-wide - shot 2 = medium - shot 3 = medium-close - Purpose: find whether moderate progression is acceptable ### Test C3 — aggressive shot delta - Workflow: text-only multi-shot - Scene: same beach, same subject - Shot progression: - shot 1 = wide - shot 2 = medium - shot 3 = close-up / angle shift - Purpose: measure where boundary disruption becomes obvious --- ## Phase 2 — anchoring strength comparison Goal: test whether stronger references materially improve continuity. ### Test R1 — text-only baseline - Workflow: text-only multi-shot - Scene concept held constant - Purpose: comparison baseline ### Test R2 — single-image anchored - Workflow: image-to-video or image-anchored omni workflow - Input: one strong reference still - Purpose: test whether one anchor reduces subject/scene drift ### Test R3 — multi-image reference - Workflow: reference-to-video / multi-image reference - Input: 2–4 consistent images of same subject - Purpose: test whether richer still references improve continuity over R2 ### Test R4 — video-reference / element-driven - Workflow: Omni with strongest available subject binding path - Input: short subject video clip if accessible - Purpose: test whether marketed Omni strength appears only with video-reference-level anchoring --- ## Phase 3 — subject type comparison Goal: identify whether continuity is easier for some subject classes than others. ### Test S1 — human subject - One person walking in stable environment - Purpose: baseline human continuity ### Test S2 — product subject - One branded object / product in stable environment - Purpose: test whether product identity locks better than humans ### Test S3 — no-character scenic shot - Pure environment / no person - Purpose: isolate camera continuity from character identity drift --- ## Phase 4 — environment complexity comparison Goal: see whether simpler scenes stabilize better. ### Test E1 — simple environment - Beach / desert road / empty studio - Purpose: easiest continuity conditions ### Test E2 — medium environment - Cafe / library / room interior - Purpose: moderate visual complexity ### Test E3 — complex environment - Busy street / crowd / many moving background elements - Purpose: determine failure boundary under production-like complexity --- ## Phase 5 — motion-vector continuity Goal: see whether directional movement continuity is a strong determinant. ### Test M1 — forward movement only - Subject keeps moving in same direction across all shots ### Test M2 — lateral tracking only - Camera tracks from same side all the way through ### Test M3 — directional conflict - Introduce subtle direction/angle contradiction on purpose - Purpose: measure how fragile continuity is to motion inconsistency --- ## Next critical experiments These experiments are explicitly designated as the next core quality-reproduction work, but deferred for later execution. - C1 — minimal shot delta continuity baseline - C2 — moderate shot delta continuity test - R2 — single-image anchored continuity test - R3 — multi-image reference continuity test - R4 — video-reference / element-driven continuity test Current status: deferred intentionally; keep as priority experiments for the next quality-focused phase. ## Priority order If budget is limited, run in this order: 1. C1 — minimal shot delta baseline 2. C2 — moderate shot delta 3. R2 — single-image anchored 4. R3 — multi-image reference 5. R4 — video-reference / element-driven 6. S3 — scenic no-character control 7. C3 — aggressive shot delta stress test --- ## Recommended data capture for every run Store all of the following: - exact payload - model name - duration - mode - query response - final video URL - local downloaded result - user judgment notes - continuity rubric scores - cost / final unit deduction --- ## Naming convention for runs Use names that make the experiment readable later. Examples: - `cont_c1_text_beach_sidefollow_min-delta_v1` - `cont_r2_single-image_beach_walk_v1` - `cont_r3_multi-image_beach_walk_v1` - `cont_s3_scenic_ocean_min-delta_v1` --- ## Decision threshold Treat a workflow as pipeline-worthy only if: - continuity_score >= 4.0 - no major identity break - no major scene reset - shot boundaries are noticeable at most mildly Anything below that stays in experimental status. --- ## Current working hypothesis Most likely route to materially better results: 1. small shot deltas 2. strong subject anchoring 3. structured chronological prompting 4. simple environments first 5. video-reference / element workflows for premium continuity goals ## Spend-control execution rule Before running any matrix row: - assume each create is billable - do not run multiple candidate variants for the same row without explicit approval - if one candidate already succeeds, stop and evaluate that output first - separate cheap documentation narrowing from live billable experimentation ## 2026-03-29 policy update For production design: - one scene should fit inside one clip (up to 15 seconds) - multi-clip chaining should no longer be treated as the primary single-scene continuity strategy - continuity testing across multiple clips is still useful for scene-to-scene transitions, not for pretending several clips are one scene