MODEL FIELD NOTES · Direction

Controlling Two-Actor AI Video Prompts: Fixed-Seed A/B/C Tests and Auto-QC False Positives

Directing two synthetic characters to approach, hold hands, or exchange a close reaction is where plain text prompts fail most often. This record documents a fixed-seed A/B/C prompt structure test and explains why automated quality scoring cannot replace 3D motion blocking and human A/B take selection.

Porncut editorial team · Published · Tested

阅读本站中文版 · 在黄果制片阅读中文版

What changed across the fixed-seed A/B/C prompt test

On September 5, 2026, we ran a controlled comparison on two-character interaction clips using a fixed random seed and identical starting conditions. Variant A used atmospheric prose describing mood and relationship tension; Variant B broke the action into explicit two-second physical beats; Variant C added left-right spatial blocking and camera framing constraints.

We archived the generated clips alongside dialogue transcripts and frame-by-frame human review notes. Beat-segmented physical verbs (Variants B and C) reliably triggered character movement where atmospheric prose left the actors nearly static. However, even the strongest text-only candidate exhibited hand drift and occlusion artifacts during close contact, confirming that prompt wording alone is not a deterministic pose controller.

When two characters occupy overlapping bounding boxes, text encoders struggle to bind specific verb phrases to the left or right actor across all twenty-four frames per second. Explicit spatial tokens reduce ambiguity, yet only a geometric skeleton rig guarantees that Actor A and Actor B maintain distinct torso trajectories throughout the shot.

What changed across the fixed-seed A/B/C prompt test
Prompt VariantStructural StrategyObserved Motion ResponseHuman Frame Audit
Variant A (Atmospheric prose)Mood adjectives and emotional intentActors remained mostly static with minor swayFail: action beat not executed
Variant B (Timed physical beats)Explicit subject-verb-object beats per 2s windowTriggered approach and hand raise, unstable contactPartial: hand occlusion artifact on approach
Variant C (Spatial + framing)Left/right actor assignment + contact target + lensClosest match to intended blockingBest of group, still required A/B take filtering

Why automated quality checks pass clips that humans immediately reject

During the same experiment, we compared an automated quality-control gate (keyframe sampling plus vision-language alignment scoring) against human inspection. Several clips that received an automatic PASS were rejected within seconds by human reviewers.

Two failure modes consistently fooled automated scoring: a hand crossing in front of a face (which causes transient facial melting or finger fusion between sampled keyframes) and extra limbs appearing during two-actor embraces. Because individual sampled frames still matched the scene description and maintained sharp background textures, the automated gate scored them high despite broken anatomy in motion.

Why automated quality checks pass clips that humans immediately reject
Failure ModeAutomated QC VerdictHuman Reviewer VerdictPorncut Studio Mitigation
Hand crossing the facePASS (high semantic and sharpness score)REJECT (transient facial distortion during occlusion)Preview camera orbit and hand trajectory in 3D Director
Extra limb on two-actor contactPASS (missed between sampled frames)REJECT (duplicate forearm during close hold)Bind pc_humanoid@1 skeleton + render A/B pair for human pick

Why Porncut pairs 3D skeleton blocking with A/B takes

These findings shaped Porncut’s core studio loop. Rather than pretending a single prompt will produce a flawless two-actor take on the first try, the workflow separates spatial planning from stochastic sampling so directors can preview blocking before spending GPU credits.

  1. Block two-actor spacing and hand IK reachability (0.045 m to 0.815 m) on the pc_humanoid@1 rig in the public 3D Motion Catalog (/motions).
  2. Assign one clear physical beat per shot across a 3-shot storyboard instead of crowding multiple actions into one clip.
  3. Generate two variants (Take A and Take B) under a fixed credit quote so a human director always chooses the clean cut.